1. INTRODUCTION

Understanding the behavior of high-impact loading processes in concrete structures is necessary to improve the impact resistance of materials used for the construction or reinforcement of several civil engineering applications. Appropriate measurement and evaluation methods for the performance analysis of various mineral-bonded composites and the validation of numerical simulations are being developed by the Research Training Group GRK 2250 of the German Research Foundation [1]. Due to the high strain rates that characterize dynamic events, devices that can acquire information at high frequencies are required. Tensile [2], compression, shear [3], impact-face [4], and impact-far response [5] of strengthened specimens under dynamic processes are also evaluated using, among other measurement techniques, high-speed cameras. Imaging devices allow not only the visualization of short-time events, but also the quantification of the material properties using photogrammetric methods. Compared to conventional measurement techniques, such as accelerometers or strain gauges, they offer both very high temporal and spatial resolution, allowing for simultaneous measurements at thousands of points.

In addition to a suitable spatio-temporal resolution to capture the material behavior of interest, the reliability of photogrammetric analysis techniques depends on the image quality. Here, the sensor technology of a high-speed camera may also affect the performance of the images obtained. High-speed cameras, such as the SMIX multichannel framing camera with a CCD matrix sensor from Specialised Imaging Ltd. in the UK, can reach speeds of up to 1 billion frames per second (fps). Because of the evident advantages of high-speed photography, the development of high-speed optical devices has been rapid. As each new camera developed comes with a combination of technologies, the descriptions used to classify such cameras quickly became obsolete. Although there are no clear boundaries for this classification, The Focal Encyclopedia of Photography [6] divides high-speed photography into four categories: High Speed: 50–500 fps; Very High Speed: 500 fps to 100 kfps; Ultrahigh Speed: 100 kfps to 10 Mfps; and Super High Speed: in excess of 10 Mfps. However, these designations are often incompatible with the terms used in the marketplace. Cameras capable of more than 100 kfps to 1.1 Mfps, such as the well-known cameras produced by Photron Limited in Tokyo, are, for instance, called high-speed cameras by the manufacturer.

Depending on the testing application, CMOS sensors are preferred over CCD sensors. The reason for this is that CMOS sensors read out the charge or voltage for each pixel directly, while CCD sensors transfer the information from pixel to pixel and into the register. This pixel-to-pixel transfer allows each photosensitive part to be larger, but substantially decreases the pixel readout rate [7]. To date, the disadvantage of high-speed CMOS sensors has been the considerable loss of spatial resolution as the temporal resolution increases. One of the latest developments is the merging of some features of CCD and CMOS technologies called an in situ storage image sensor (ISIS) [8]. This technology integrates a memory in the pixel itself or incorporates features of the CCD readout processes, thus increasing the data transfer rate and, depending on the sensor architecture, the light sensitive part of the detector. In this way, higher frame rates can be achieved without losing dimensional resolution.

A. State of the Art

For an adequate performance of concrete structures under dynamic loading, the progressive formation of fine cracks under increasing tensile loading is aimed at to ensure high nonelastic deformation before reaching tensile strength [1]. In [9], a method to determine crack propagation velocities from high-speed image sequences recorded with a Photron Fastcam SA-X2 480K is presented. Crack velocities are obtained as derivatives of crack lengths over time using a five-point least-squares method to also obtain standard deviations as internal quality measures. The results show that the crack velocity could only be obtained for a maximum of five time steps. This is due to the set camera configurations as a trade-off between the spatio-temporal resolution and the precision of the configuration system. A frame rate of 160 kfps and an image format size of $256 \times 184\;{\rm px}$ are configured. Although the evolution of the crack tip propagation is clearly recognizable in the image sequences, the precision of the crack propagation velocity might be improved by increasing the image format size and the number of frames before the crack reaches the surface mesh boundaries. As investigated in [10], the Photron Fastcam SA-X2 480K presents a high measurement performance using appropriate testing configurations to capture highly dynamic events. However, its technical characteristics [11] do not allow measurements at the highest frame rates with an acceptable width/height image ratio: The maximum possible spatial resolution is $128 \times 48$ pixels (px) at 480 kfps.

In addition to crack deformation analysis, the determination of the velocity of the stress wave introduced by the impact is also of interest. The distance traveled by the wave through the specimen particles in a given time interval causes the particles to acquire their own velocity and thus to deform under the impact action [12]. To avoid significant influence of crack propagation forces, data from the deformed specimen before the main crack formation is sought. This is intended to improve the reliability of estimating mechanical properties in highly dynamic material tests, such as the wave velocity, in future investigations.

To increase the dimensional and temporal resolution of image sequence data before and after main crack initiation, the measuring potential of three cameras belonging to the category of ultrahigh-speed cameras according to [6] is presented. The cameras tested with the maximum frame rate are the Phantom TMX-7510 with 1.75 Mfps (P1.75M), the Shimadzu HPV-X2 with 10 Mfps (S10M), and the Kirana7M with 7 Mfps (K7M). These ultrahigh-speed cameras offer options to record, analyze, and understand impact events, such as those conducted in civil engineering material testing, since different stages of a deformation process can be visualized in more detail. A prerequisite for these image sequence-based analyses is a thorough understanding of the quality of the image data and the precision potential of geometric measurements derived from image sequences. For this reason, the potential of the cameras is compared regarding the measurement accuracy and validated in real concrete impact applications. The latter aims to gain an understanding of the precision of the crack velocity computation and to test the possibility of reliably measuring specimen deformation prior to main crack propagation. For the analyses prior to the main deformation, the displacement of the specimen is obtained for each point of a mesh surface. The direction of the motion of the mesh surface points over time is referred to as surface particle propagation in this work.

The application potential of the three cameras is discussed in [13] for the S10M and in [14] and [15] for earlier models of the Kirana and Phantom, respectively. In [13], a stereo high-speed S10M system has been used to extract the 3D position and time information to quantify properties as fragmentation severity, debris velocity, and debris spread of a hypervelocity impact experiment in 24 µs. An 82% success rate of the fragment trajectories were tracked based on the crater information. The Kirana5M is used to visualize the evolution process of the maximum principal strain under the action of blasting before stress wave reflection (or significant crack formation) during 100 µs [14]. The results show a strain partitioning of the “elastic vibration area” into the “plastic area” and “quasielastic area.” The Phantom V711 is used to observe the interaction of dispersed liquids and shock waves that cause droplet atomization, which can lead to risk scenarios in the presence of hazardous materials [15]. Shock waves could be clearly observed at a time step of 45 µs, but the size of secondary droplets after atomization could not be detected due to the insufficient pixel resolution.

To describe the potential of the three cameras, this work is divided into the following sections: First, the selected experimental setups of the three-point bending (3PB) test and split-Hopkinson pressure bar (SHPB) test are introduced. Second, a characterization of the cameras is given to understand the performance of the sensor technology and the possible image quality challenges. Third, the results of real photogrammetric applications, such as the deformation analysis on the surfaces of concrete specimens, are presented. This includes the propagation of surface particles up to the main crack formation and the crack propagation velocity. Finally, the conclusions of the sensor performance and the outlook motivating this work are discussed. The complete experimental data and results are available upon request from the author responsible for correspondence.

2. EXPERIMENTAL SETUPS

The structural deformation processes of surface particle propagation before main crack formation, followed by crack propagation, are analyzed. Figures 1 and 2 show the selected experimental setups: Drop-weight facility for the 3PB test in $160 \times 40 \times 40\;{{\rm mm}^3}$ cuboids, and SHPB for the compression test in 80 mm long and $50\;\rm mm\,\emptyset$ diameter cylinders. 2D spatio-temporally resolved measurements are obtained from the monocular camera systems presented.

 

Fig. 1. 3PB tests recorded with the three monocular optical systems. For each camera, the image size at different frame rates is shown, as well as the two-wire circuit for triggering the K7M and the S10M.

Download Full Size | PDF

 

Fig. 2. SHPB tests recorded with a K7M monocular optical system. The camera field of view, triggering system, and impactor facility characteristics are also shown.

Download Full Size | PDF

Small-scale specimens are used to observe the deformation behavior of the structure under load. Here, the relative displacements are expected to be small, so a high spatial resolution of the sensor is required. For this reason, only the main deformation region of the specimens is observed with the three cameras in the 3PB experiments. As presented in Section 4, surface particle propagation analysis is carried out for both experimental setups. However, crack propagation velocities are only presented for the 3PB tests. The reason for this is that the SHPB tests are only recorded with the K7M and there are no crack propagation data from the other cameras to evaluate the plausibility of the resulting velocities. The 3PB experimental setup is chosen because the surface particles and the cracks can propagate in-plane in either the horizontal or vertical direction. SHPB is selected because the length of the incident and transmitted bar of this experimental setup must be sufficiently long to ensure one-dimensional wave propagation [12]. Therefore, it is expected that the main direction of particle propagation will be in the direction of the impact and that the analyses before crack formation are not significantly affected by out-of-plane errors. In addition, these tests are chosen for their ease of repeatability because no glue is needed to attach the specimen to the experimental facility.

As shown in Fig. 1, 3PB is recorded with the three cameras. It can be seen that the image size of the K7M and the S10M does not vary with the frame rate. In contrast, the image size of P1.75M at full resolution ($1280 \times 800\;{\rm px}$ at 76 kfps) decreases significantly at higher frame rates. In addition, the sensor size of the K7M allows the observation of the impactor without considerable loss in the dimensional resolution of the specimen for both tests. This is convenient because the response of the specimen under the impact effects can be analyzed simultaneously in the time domain. However, due to the low light sensitivity of the K7M, a flash lamp is required that fires after the camera is triggered. According to [16], the flash lamp takes 80 µs (30 µs flash delay ${+}{50}\;\unicode{x00B5}{\rm a}$ rise time) to reach full brightness. When the light reaches this peak, it remains constant for 1 ms. This system has a light duration of 2 ms and uses a U-shape Xenon flashtube as the light source. The P1.75M and S10M, on the other hand, do not require such a large amount of light and can be operated with conventional lamps. The 1kW-LED-4438 lighting system with 100,000 lm in continuous operation was used for both cameras. However, the GOM LED light projector for the P1.75M and the ARRI Pocket par 400 for the S10M in continuous operation also proved sufficient to provide good lighting. The former uses blue light with a power consumption of about 80 W and the latter uses white light with 400 W.

Camera triggering is an important aspect because the deformation processes of surface particle propagation and crack propagation velocities can be properly evaluated before any noticeable damage to the concrete occurs. Depending on the recording capacity of the camera, a higher precision of the triggering time is relevant. Unlike the K7M and the S10M with ISIS technology, which only allows the storage of extremely short image sequences, the P1.75M with CMOS technology can generate a large amount of data at various frame rates and image formats. Because of this large recording capacity, the P1.75M can even be manually triggered to capture the 3PB tests. On the other hand, the first two cameras are triggered by closing a two-wire circuit at the moment of impact, as shown in Fig. 1. By triggering at the moment of impact, it can be difficult to obtain images prior to the main deformation. This is particularly important for the K7M recordings, since the 80 µs to full brightness of the flash lamps must be considered. However, images can also be obtained here with the circular buffer storage capability of these cameras. Therefore, sufficient images were obtained for both particle and crack propagation analysis.

At the SHPB facility in Fig. 2, a strain gauge located on the input bar and connected to the oscilloscope PicoScope4000Series is used for triggering. The triggering process is shown in Fig. 3, where a frame rate of 200 kfps is used to observe the entire experiment.

 

Fig. 3. K7M triggering process showing the input and output strain signals of a SHPB test recorded at 200 kfps to capture the entire experiment. A triggering signal delay of 61.547 µs and a configured camera delay of 70 µs are plotted. Changes in length ($\Delta LX$) are used to define the wave crossing time to select a suitable frame rate, as presented in Fig. 10.

Download Full Size | PDF

Based on a rising strain gauge signal, a threshold can be set in the trigger controls of the oscilloscope’s software. When the signal exceeds this threshold, the software sets this point as time $t = 0$ and sends the trigger signal to the camera. The camera receives this signal with a consistent trigger delay average of 61.559 µs and standard deviation of 0.012 µs obtained from 14 experiments. In addition, a delay was configured in the camera’s software to allow the flash lamps to reach full brightness, but also to capture the time of interest. After the impactor is loaded with 1.5 bar, a compressive wave (green curve) is generated and propagates toward the specimen. When the wave arrives at the interface between the input bar and the specimen, a part of the wave is reflected back into the input bar toward the impactor end and the rest transmits through the specimen into the output bar (red curve) [17]. The K7M is then triggered (gray curve), when the threshold condition using the input strain gauge signal is satisfied at the time of 0 ms. Due to the aforementioned signal trigger delay, the camera is triggered 61.547 µs later. Recording starts after a configured camera delay of 70 µs.

The changes in length in the horizontal direction ($\Delta LX$) between input bar–cylinder, in-cylinder, and cylinder–output bar can be observed in the lower diagram of Fig. 3. When the change in length between cylinder–output bar begins to decrease, it is assumed that the compressive wave has reached the output bar and thus has passed through the specimen. During this time (blue filling), the propagation of the surface particles can be observed. After this time, the main specimen deformation begins, and the crack propagation velocity can be evaluated. This analysis allows the selection of an appropriate frame rate, high enough to observe the small relative surface deformations, but also long enough for the K7M’s recording capacity to capture both deformation processes.

3. CAMERA CHARACTERIZATION

As mentioned in the introduction, the reliability of the photogrammetric analysis is closely related to the image quality. Although the experimenter has a significant influence on the image performance through the configuration of the optical system, the challenges to overcome the image shortcomings can be explained by the sensor technology. This section provides an overview of the sensor operation to understand the image acquisition challenges discussed below. Relevant technical features that characterize the sensor performance of the three cameras are presented in Table 1.

Table 1. Technical Information for the Tested High-Frequency Camera Sensors^a

View Table | View all tables in this article

A. Sensor Operation

The sensor of the P1.75M works like a conventional CMOS sensor, but instead of using front side illumination (FSI) like most high-speed cameras, it uses back side illumination (BSI) [21]. This results in a fill factor of nearly 100% and an increase in the processing speed to achieve higher frame rates. Processing speed is increased by replacing the circuitry that normally conducts light to the photodiodes with other metal wiring to reduce the resistance at high frame rates. As a result, this camera can capture up to 1.75 Mfps at a spatial resolution of $1280 \times 32\;\rm px$. At this frequency, the width/height image ratio can be improved to $640 \times 64\;\rm px$ by using the binned mode. In this binning mode, four pixels ($2 \times 2$ squares) are grouped into one large pixel to increase the throughput and the light sensitivity, but the image resolution is also reduced. A final important advantage of this camera is the ability to store image sequences of up to 2.3 s at the highest frame rates. This is due to the large number of analog-to-digital converters integrated in the CMOS sensor to generate a large amount of data. Therefore, the precise timing of a trigger system is largely irrelevant here.

Although the fill factor and the processing speed are improved with this BSI technology, the CMOS operation of the P1.75M does not allow high image frequency without sacrificing spatial resolution. To increase both resolutions, the S10M and K7M ultrahigh-speed cameras use hybrid sensors (a combination of CMOS and CCD features). In the S10M, global shutter operation is performed simultaneously in each pixel to convert photoelectrons into voltage and reduce noise (CMOS technology), and then the signal is read out from the pixels to the on-chip memories (FTCCD: Frame transfer CCD). The ultrahigh-speed CMOS (uCMOS) sensor of the K7M integrates CMOS transistors for reset, readout, and pixel selection as well as CCD memory cells for storage in each pixel.

The sensor functionality of the S10M is described in [23]. The FTCMOS sensor consists of two spatially separated regions: a pixel region and a memory region. The pixel region consists of a pinned photodiode (PD), a floating diffusion (FD) for charge-to-voltage conversion, in-pixel noise reduction, in-pixel source-follower current sources (SF), multiple pixel output wires, and pixel select switches. The in-pixel SFs ensure that the current is large enough to prevent a decrease in the speed of the signal readout from the pixel to the on-chip memory. The in-pixel noise-reduction circuit allows the number of readout signals from the pixel to the on-chip memory to be reduced, since no additional memory is required to read out the reset noise of the FD. Reducing the number of readout signals increases the recording capacity, improves the readout speed, and reduces the power consumption. The use of multiple pixel output wires to read out the signals from four pixels to the on-chip memory via a single output wire also improves the readout speed and enables flexible pixel switch selection to increase the frame rate and recording capacity. The memory region is formed by 128 on-chip memories connected to each pixel through the output wires. This feature also enables high-speed recording, but it limits the maximum image sequence length to 128 frames when the complete image format is being used in full pixel mode (FP) at a maximum frame rate of 5 Mfps. An increase in the recording capacity to 256 frames and in the frame rate to 10 Mfps can be obtained using the pixel select switches in half pixel mode (HP) operation. In the HP mode, half of the pixels are selected, and the signal of each second pixel in a diagonal pattern is stored. Finally, a light-shielding layer is placed on the on-chip memory to prevent distortion of the stored signals.

The structure of the K7M sensor is described in [24]. The architecture of each pixel is divided as follows: A PD, a two-dimensional memory bank (MB) of 180 memory cells ($18 \times 10$, thus limiting image sequence length to 180 images), the PD to CCD MB input structure, the MB to FD output structure, and a pixel readout circuit. This architecture allows correlated double sampling for low-noise operation. The MB operates in three-phase operation (each cell unit has three separate polysilicon electrodes) to move the charge around within the in-pixel CCD cells using vertical and horizontal transfers. Charge transfer efficiency has been optimized from the PD to the readout register, resulting in short transfer times for high recording capacity. In addition, the memory cells are protected by light shielding that minimizes contamination of the stored images by incident light. The architecture of this sensor is characterized by scalable resolution, since the memory cells are located within each pixel, and scalable recording capacity, as a trade-off with pixel size.

B. Image Acquisition Challenges

The following challenges have been identified based on the operation of the sensors of the three cameras along with the technical information presented in Table 1 and the corresponding images captured during the 3PB and the SHPB tests:

1. Recording Capacity

Considerations of maximum image sequence lengths are particularly relevant for the ISIS sensors presented in this work because their operating technology rather strictly limits the expansion of the number of memory cells. Although the pixel architecture of the K7M allows a scalable memory depth, the 180 memory cells (number of frames) must fit into each pixel with the PD. Increasing the number of memory cells would further reduce the effective imaging area. The recording capacity of the S10M depends on the operating mode used. The FP mode allows 128 frames to be stored, which corresponds to the number of on-chip memories. To obtain a higher recording capacity, the HP mode can be selected to store 256 frames. As mentioned before, in the HP mode, half of the pixels are selected and therefore, the image resolution is reduced by half. Despite the resulting image size is the same as it would be in the FP mode, this reduction affects the quality of the image correlation process.

Although the 128 frames of the S10M and the 180 frames of the K7M are sufficient to recognize high-speed dynamic events, the recording time remains limited; to benefit from the full recording capacity, a very precise and reliable trigger system is required. In the case of the P1.75M, the recording capacity is 2.3 s at the highest temporal resolutions. Therefore, considering that external triggering is advantageous for experimental analysis, a simple manual trigger can be used to capture the fast event.

2. Sensor Sensitivity

For high-speed imaging, the sensitivity of the sensor is an important aspect due to the lack of exposure time, which is of course limited by the frame rate. The P1.75M’s BSI technology, which improves the fill factor and dynamic range, gives it the highest light sensitivity at ISO 40,000 in “standard” mode and ISO 50,000 in “binned” mode. With respect to the S10M, the reported sensitivity at ISO 16,000 is largely due to the fill factor, but also to the full well capacity, photon-electron noise, and conversion gain. These technical features also influence the light sensitivity of the K7M, but no data is provided by the manufacturer.

The S10M reaches a fill factor of 55% because the memory region is separated from the pixel region, allowing a larger PD size. In contrast, even though the microlens improves the K7M’s fill factor, this camera has a relatively low value of 11%. The reason for this is the space needed for the in-pixel memory cells, which results in a small photodiode size of $5 \times 28\;{\unicode{x00B5}{\rm m}}$. In addition, although the full well capacity of the K7M is higher, the lower noise of the S10M results in a wider dynamic range. Consequently, the spectral response of the S10M is extended into the UV light. Finally, the higher conversion gain of S10M provides better performance in low-light conditions. However, a high gain value may eventually result in a degradation of the signal-to-noise ratio (SNR) [7].

Due to the high light sensitivity of the P1.75M and the S10M, conventional lamps provide sufficient light for the high-speed imaging of the experimental setups presented in Section 2. In contrast, the K7M requires a flash lamp that takes some time to reach full brightness. However, the variations in light intensity inherent to flash lamps affect the image information and consequently also the image sequence correlation process. Therefore, the challenge is to consider the light behavior of the flash lamps when setting up the configuration system and in the photogrammetric process.

To compare the imaging performance of the three cameras, a rigid-body motion-based procedure is performed on static recordings acquired at different frame rates, as proposed in [10]. As 2D measurements are obtained from monocular image sequences, plane similarity transformations between two consecutive time steps are performed to define plane mapping shape discrepancies. These discrepancies are plane mapping errors (${s_{\!0}}$), which result from an apparent variation in the reconstructed plane surfaces over the acquisition time, including translation, rotation, and scale, despite that no motion is presented. Since the camera-to-specimen distance does not vary, the scale parameter is set as being equal to 1. Although estimating the parameters would require only two identical points, the transformation parameters can be calculated by an overdetermined least-squares adjustment [25] for the sake of redundancy. The resulting ${s_0}$ are shown in the three boxplots for each camera in Fig. 4.

 

Fig. 4. Boxplots of plane mapping errors (${s_0}$) obtained from statically recorded specimens with image size in px used to analyze the imaging performance of the three cameras.

Download Full Size | PDF

The higher image quality of the P1.75M is demonstrated by the significantly lower ${s_0}$ errors compared to the other two cameras. However, the significant loss of spatial resolution with increasing temporal resolution can also be observed in the given image sizes in px. The ${s_0}$ of the K7M are higher than those of the S10M and show significantly more variation. The errors ${s_0}$ shown in the S10M boxplot are higher in the HP mode than in the FP mode images taken at the same frame rate. This shows the effect of the loss of spatial resolution on photogrammetric object reconstruction. Finally, a decrease in the SNR with an increasing frame rate can be observed for all three cameras. For the S10M, a noticeable decrease in the SNR is only observed at 10 Mfps, which is only possible in the HP mode.

3. Image Lag

Transferring all of the photogenerated electrons from the PD to the FD can be an issue in high-speed applications because of insufficient charge transfer results in incomplete charge depletion. The result of incomplete charge transfer is image lag, also known as the ghosting effect. It retains information from one frame to the next frame [26]. There are different strategies to reduce image lag [27]: On the one hand, the PD fabrication and transfer gate operation can be optimized to avoid residual electrons in the PD region to be transferred in the subsequent frames. On the other hand, increasing the charge transfer efficiency (CTE) from the PD to the FD is a critical parameter since the CTE signal is strongly related to the transfer voltage. Due to lower power supply voltage that characterizes CMOS sensors, a low-voltage transfer pulse can result in incomplete charge transfer.

In the K7M, the charge packets are transferred many times through the CCD–MB array before they reach the FD. For high-performance purposes, the pixel structure is designed to maximize the fringing field that transports the charge from one memory to the next in the CCD array, and to optimize the control of the charge under the memory gates using an effective low-voltage channel. An appropriate design is important because imperfections in the silicon-based interfaces create electrically active traps in the path of the moving charge packet. Some of the trapped charges can be released asynchronously to the charge transfer, causing image lag [24].

The charge transfer process from the PD to the FD in the S10M takes place in the pixel region. An optimization to generate a horizontal electrical field to drift the photoelectrons by changing the doping concentrations in the pixel fabrication is reported in [23]. After in-pixel noise reduction, the signal from four pixels is transported through a shared output wire to the corresponding on-chip memories in the memory region.

In CMOS matrix sensors, such as the one used in P1.75M, the charge generated by the incident light is directly processed in each pixel element. There is therefore no sequential charge transfer. Consequently, individual sensor elements can be addressed or processed, and there is a lower sensitivity to blooming and transfer loss [25].

To observe the effect of image lag, Fig. 5 shows the image brightness (${B_I}$) variations of the frames recorded for the static recordings of the three cameras presented in Fig. 4. ${B_I}$ is defined in [25] as

 

Fig. 5. Brightness (${B_I}$) variations, as defined in Eq. (1), of the static recordings evaluated in Fig. 4 for imaging performance analysis.

Download Full Size | PDF

Insufficient charge transfer by the K7M is clearly seen in the pattern behavior every 10 frames, which corresponds to the number of vertical rows in the MB. Image lag affects every frame but is more apparent when the intensity varies significantly between frames. As explained in Section 2, the flash lamps are triggered after the camera is triggered. A camera delay is set to allow the lamp to reach full brightness. During the delay time, the sensor is already exposed to light, so the first frames are more affected by this image lag phenomenon. Finally, the brightness of frame 180 is higher than the previous frames. The most likely reason for this is that the leftover charge is depleted to the readout circuitry with the last frame. The higher brightness of these frames (10 first and last) can be minimized by using the “Capture Mode” of “Delayed Stop,” as reported in Table 1. In this mode, when the camera is ready to be triggered, it begins capturing at the specified frame rate. When the camera is then triggered, the flashes can be programmed to start a few nanoseconds after the trigger, while the camera continues capturing frames for a further specified camera delay. A configured number of frames before and after the camera delay are then stored. In this way, due to the camera’s storage process, it will be possible to avoid the first frames captured, which would normally be more affected by the brightness of the flash lamps.

This image lag effect of previous frames overlaying the subsequent frames of a Kolsky bar experiment recorded with a Kirana camera is reported in [28]. It is concluded that the ghosting effect is particularly pronounced with significant object translation and it becomes more apparent with high contrast. Therefore, contrary to the white–black contrast preferred for image correlation processes, a gray toner background is better suited for the Kirana cameras.

Unlike the brightness response of the K7M, the brightness response of the S10M and P1.75M remains constant over recorded frames with no apparent image lag. The direct photo-electron conversion from the PD to the FD, without additional charge transfer, is most likely the reason. However, the brightness of the P1.75M is more constant than the brightness of the S10M because of the superior performance of the BSI pixel technology.

 

Fig. 6. Fixed pattern noise of a K7M image section sequence without applying the Session Black correction. A red circle around an apparently stationary pixel illustrates this effect.

Download Full Size | PDF

4. Fixed Pattern Noise

This type of noise is caused by variations in the semiconductor of the CMOS, which leads to radiometric distortions in the image. Three major contributions for each pixel are explained in [29]: the offset noise caused by semiconductor variations within the PD and the amplifiers, the dark noise caused by leakage current and a pixel will be charged even without light, and the photo response non-uniformity noise due to slightly different sensitivities of each pixel.

Since these noise models are well understood by the manufacturers, digital compensations to correct the spatial nonuniformity of the intensities observed across the sensor are possible. Therefore, most commercial high-speed cameras are equipped with a flat-field correction to provide a black reference for all pixels in the sensor, as shown in Table 1. This correction is not available in the S10M because the in-pixel noise reduction circuitry eliminates the fixed pattern noise before reading the signals from the pixel region to the memory region [23]. For the P1.75M and K7M, the correction should be performed before each recording. The effect of not performing this correction is more visible at high light levels. For this reason, this effect is particularly evident in the K7M images taken without prior black correction, as can be seen in Fig. 6. It can be observed that there are some apparently stationary pixels (for example, the pixel surrounded by a red circle) that do not change their coordinate pixel position in the recorded images, even though the cracks in the specimen are propagating and opening.

4. HIGH-IMPACT LOAD APPLICATIONS

To interpret the experimental results and to characterize the properties of the material, the specimens should deform uniformly under controlled testing conditions. In quasi-static experiments, the testing conditions are monitored and adjusted in real time so that the specimen deforms under specified conditions throughout the test. In impact experiments, the test conditions cannot be controlled and the testing conditions on the specimen depend on the specimen’s response [17]. The specimen response is observed using the three cameras: the P1.75M, the S10M, and the K7M. Their potential is evaluated by comparing the ability of the sensors to generate data to observe the spatio-temporal surface particle propagation and crack propagation velocities in 3PB tests. The potential of the K7M to generate such deformation data is additionally evaluated in SHPB tests. However, in the absence of data recorded by the other two cameras to compare crack propagation velocity results, only surface particle propagation analyses are presented here.

For the 3PB tests, the particles are expected to deform gradually from the top of the surface where the specimen is impacted to the bottom of the surface. At the bottom, the main crack initiates and propagates through the specimen to the region where the bending deformation is higher. For the SHPB tests, the particle deformation is expected to occur mainly in the impact direction due to the one-dimensional induced wave and, as observed from the conducted experiments, the main crack initiation occurs in the impact side region where the deformation is higher. Since the main direction of the presented deformation processes occurs in-plane, using a stereo camera system to obtain 3D measurements is not required to capture the evolution. The deformation analysis for 3PB and SHPB with the best performing results are presented below.

A. Surface Particle Propagation

As mentioned in the introduction, the determination of the velocity of the stress wave introduced by the impact that causes the deformation of the tested specimen is of interest in material testing analysis. The wave velocity can be derived from the motion of mesh surface points, which represent the deformation of the specimen particles caused by the induced wave. According to the one-dimensional wave theory in solids as presented in [12], the velocity of the propagating particles ($v$), defined as the variation of the displacement in a given time interval ($\frac{{\delta u}}{{\delta t}}$), could be used not only to calculate the strain ($\varepsilon$) at each specimen surface position, but also to estimate the velocity of the stress wave ${c_L}$ traveling through the specimen, as

Data prior to main crack formation is sought to avoid significant influence of crack formation on surface particle displacement for wave velocity determination. To do this, the ability to capture the particle propagation is tested in 3PB using all three cameras and in SHPB using the K7M. The results are presented in Figs. 7–10. For each camera test, the particle propagation in terms of displacement and the main crack formation in terms of strain are acquired and visualized using the GOM Correlate Pro 2022 Service Pack 3 software from Carl Zeiss GOM Metrology GmbH. As shown in [2], the strains can be used to represent the crack fracture process of a specimen and are therefore used here to represent the crack formation. For 3PB, the in-plane displacements $\parallel \!\vec u\!\parallel$ provide a better representation of the particle propagation than in the impact direction. In addition, the horizontal strains ${\varepsilon _X}$ provide clearer results to represent crack formation due to the impact bending condition. Since the strains over the surface do not unambiguously indicate the crack formation in the early stages, the displacements in the orthogonal direction of the crack ${u_X}$ provide a clearer definition of the main crack initiation. They are used to observe at which time step the particle propagation motion is mainly influenced by the horizontal direction caused by the crack formation.

 

Fig. 7. $\parallel \!\vec u\!\parallel$ particle propagation, and ${u_X}$ displacements and ${\varepsilon _X}$ strain to define main crack initiation and formation for 3PB recorded with the P1.75M.

Download Full Size | PDF

 

Fig. 8. $\parallel \!\vec u\!\parallel$ particle propagation, and ${u_X}$ displacements and ${\varepsilon _X}$ strain to define main crack initiation and formation for 3PB recorded with the K7M.

Download Full Size | PDF

 

Fig. 9. $\parallel \!\vec u\!\parallel$ particle propagation, and ${u_X}$ displacements and ${\varepsilon _X}$ strain to define main crack initiation and formation for 3PB recorded with the S10M.

Download Full Size | PDF

 

Fig. 10. ${u_X}$ particle propagation, and ${u_Y}$ displacements and $\varepsilon$ principal strain to define main crack initiation and formation for SHPB recorded with the K7M.

Download Full Size | PDF

The 3PB results in Fig. 7 show the superior image quality previously reported for the P1.75M in Section 3.B, which produces less noisy results even when only a region of $6\;{{\rm cm}^2}$ (${\approx} 10 \times 0.6\;{{\rm cm}^2}$) of the full surface ($4 \times 16\;{{\rm cm}^2}$) is analyzed using approximately $640 \times 59\;{\rm px}$. The increase of the particle propagation displacement can be observed smoothly up to $64.70\;\unicode{x00B5}{\rm s}$. At this time step, the high strain values (0.15%) and the lateral surface displacements (${-}24.1 \lt {u_X} \lt 20\;{\unicode{x00B5}{\rm m}}$) already indicate where the main crack will form. Therefore, the crack initiation position at the next time step (at $65.29\;\unicode{x00B5}{\rm s}$) can be clearly defined in the lower mid-surface region. The propagated crack at $75.29\;\unicode{x00B5}{\rm s}$ shows an expected strain gradient in the horizontal direction from the more deformed region (1.05–1.31%) to the less deformed surface region (${-}0.09$ to 0.15%).

The surface propagation results for the K7M in Fig. 8 and for the S10M in Fig. 9 are less homogeneous. However, although the K7M surface appears to have more distributed noise, the much higher spatial resolution of this camera allows observation of not only a larger and more detailed region of the specimen surface ($25\;{{\rm cm}^2} \approx 6.4 \times 3.9\;{{\rm cm}^2}$ with ca. $908 \times 543\;{\rm px}$), but also of the impactor. Due to the high spatial resolution, the spreading of the particle propagation from the impact region to the entire surface can be clearly observed up to 54 µs. Similarly, the lateral splitting of ${u_X}$ shows the position of the main crack, whose initiation can be well identified at the highest strain surface values three time steps later (at 56 µs). The formed crack is well visible at 100 µs, but the strains distributed over the surface emphasize the image noise.

Because of the smaller sensor resolution of the S10M, the specimen region area of $23\;{{\rm cm}^2} \approx 6.2 \times 3.7\;{{\rm cm}^2}$ is covered with about $400 \times 235\;{\rm px}$. This is less than a quarter of the pixels that are used by the K7M to capture the specimen surface. The spreading of particle displacement is visible, but the S10M propagation results are less continuous than those observed with the K7M. In addition, although ${u_X}$ and ${\varepsilon _X}$ indicate the region where the crack will appear at 107 µs, the origin of the main crack is only slightly observed 10 time steps later (at 117 µs). In addition to the lower spatial resolution due to the HP mode used for the test recording, the slower frame rate may also contribute to these results.

Comparing the three 3PB particle propagation results, it can be seen that the $\parallel \!\vec u\!\parallel$ ranges of the K7M–$[0,80]\;{\unicode{x00B5}{\rm m}}$ and the S10M–$[0,95.2]\;{\unicode{x00B5}{\rm m}}$ are noticeably larger than those of the P1.75M–$[0,49.2]\;{\unicode{x00B5}{\rm m}}$. This happens because the most deformed region before the main crack starts to propagate is not visible in the narrow field of view of the P1.75M. Therefore, the data of the maximum deformation at the time of the crack initiation is not available and could affect the wave velocity estimation.

For the SHPB results in Fig. 10, due to the cylindrical shape of the specimen, only the ${u_X}$ displacements are used to characterize particle propagation. However, main crack formation is shown using the principal strain $\varepsilon$, which indicates the largest strain in either the ${\varepsilon _X}$ or ${\varepsilon _Y}$ direction. This is due to the crack width opening, mainly vertically. Although the high $\varepsilon$ values show the crack propagation, the out-of-plane errors in $Y$ must be considered when interpreting the material properties [30]. However, the surface particle propagation in the impact direction (${u_X}$) can be considered to analyze the particle deformation caused by the uniaxial compressive wave.

In a recording time of 36.67 µs, the surface particles propagate in a range of $[0,154]\;{\unicode{x00B5}{\rm m}}$. At this time, the ${u_Y}$ displacements orthogonal to the crack direction and the highly strained surface regions indicate the crack position and therefore the main crack initiation can be clearly detected one time step later (at 37.33 µs). Compared to the results obtained in the 3PB tests, the K7M can capture these displacement values due to its sensor size and hence favorable spatial resolution, although the surface deformation prior to the main crack formation occurs in a shorter time period. In this test, a cylinder surface length of 8 cm is covered with approximately $745\;{\rm px}$. In addition, it can be observed that the displacement and strain surfaces have less distributed noise than in the 3PB case. As shown in Fig. 3, this happens because the trigger is sent to the camera when the wave crosses the input bar, giving the flash lamps time to level out the light intensity. Thus, the charge transfer performance of the K7M’s MB is less affected by the high intensity values.

Since the SHPB facility is designed to ensure one-dimensional wave propagation, an initial estimation of the wave propagation velocity ${c_L}$ prior to main crack formation can be performed according to Eq. (2). Using the deformation data in the impact direction of all the surface particles up to the time step of 36.67 µs, ${c_L}$ values between $(2200{,}3500)\;{\rm m/s}$ are obtained. To compare this wave velocity with a traditional measurement method, an ultrasonic wave measuring device is used to generate, receive, and measure the time for an ultrasonic wave to propagate through three specimen tests as presented in [31]. A digital caliper is used to measure an average cylinder length of 80.74 mm. The three tests result in a wave travel time of 17.9 µs; hence, a wave propagation velocity of ${c_L} = 4510.6\;{\rm m/s}$ is obtained. It is noticeable that the resulting wave velocities from the optical measurements and the ultrasound device differ. However, it should be considered that this initial image sequence-based estimation includes the high and low deformation values over the entire mesh surface. This high spatio-temporal information, which allows the estimation of the wave propagation velocity during the specimen’s deformation, and a proper analysis of the material behavior may contribute to improving the precision of the wave velocity estimation in future investigations.

B. Crack Propagation Velocity

The detailed explanation of the fully developed method to determine crack propagation velocities using high-speed camera image sequences is presented in [9]. An image sequence capturing the speckle pattern on the specimen’s surface is processed to obtain surface point coordinates between consecutive time steps. The surface points are triangulated into a mesh, and discontinuities in the displacement vector fields of all triangles in the mesh are used as representations of cracks.

To obtain high-quality triangulated meshes from ultrahigh-speed camera image sequences, an interest operator is used to detect the surface points with good image contrast, and subpixel accuracy least-squares image matching is applied to track the surface points [32]. After the zero-load epoch, discontinuities in the displacement vector field are computed using relative translation vectors ${\vec t_{{\rm rel}}}$ for each triangle. These are interpreted as crack openings and used as deformation vectors. The norm of the relative translation vectors $\parallel\! {\vec t_{{\rm rel}}}\!\parallel$ and a threshold value $\delta$ in px given by the precision of the displacement field are used as deformation quantities. For this, the triangles are cataloged in [9] as crack candidates if $\parallel {\vec t_{{\rm rel}}}\parallel \gt {\delta _1}$. To connect the deformed triangles of a given crack, region-growing techniques are employed using a second threshold, where ${\delta _2} = 0.5 \cdot {\delta _1}$. For the calculation of the crack length, the crack tip is automatically determined using principal component analysis. Finally, the crack propagation velocity is calculated as the derivative of the crack length over time. The mathematical model used is

For the 3PB tests presented in this work, the results of the crack propagation velocities for image sequences recorded by the three cameras are plotted in Fig. 11. Depending on the precision of the displacement field, a different threshold $\delta$ yielded the best results for each camera. The error bars in the velocity curves represent the standard deviations or precision of the estimated velocities. In addition, the camera configurations and the raw images of the specimens before crack initiation and the last tracked crack length are shown. Furthermore, the evaluated image sequences for different time steps alternately show the relative translation vector lengths $\parallel {\vec t_{{\rm rel}}}\parallel$ for each triangle using color-coded maps and the connected triangles of the crack (blue) with the crack tips (green) and the mesh borders (yellow) for the corresponding $\delta$. Statistics of the average velocities ($\bar m$) and of the standard deviations (${\sigma _m}$) are presented in Table 2.

 

Fig. 11. Crack propagation velocities of 3PB tests recorded with the three cameras and computed with the presented threshold $\delta$. The raw images before crack initiation and the last tracked crack length (manually highlighted in fuchsia) are also shown. Furthermore, image sequences for different time steps alternately show the relative translation vector lengths $\parallel {\vec t_{{\rm rel}}}\parallel$ using color-coded maps and the crack triangles (blue) with the crack tips (green) and the mesh borders (yellow).

Download Full Size | PDF

Table 2. Averaged Crack Velocities ($\bar m$) and the Corresponding Standard Deviation Statistics^a in $m/s$ for the Three Cameras^b

View Table | View all tables in this article

The crack velocity plots of the three cameras show initial velocities less than 2000 m/s. The velocity curve of the P1.75M tends to decrease with time and an average velocity $\bar m = 603.08\;{\rm m/s}$ is obtained. In contrast, the velocity curves of the K7M and of the S10M vary in a similar range over the evaluated time, resulting in $\bar m = 789.22\;{\rm m/s}$ and $\bar m = 854.26\;{\rm m/s}$, respectively. Note that the velocity of the P1.75M is observed over a longer time period of 40 µs due to the lower frame rate. In comparison, the velocity of the K7M is estimated for a time period of 19 µs and for the S10M for a time period of 11.5 µs.

The lower thresholds of the P1.75M with $\delta = 0.12\;{\rm px}$ and of the S10M with $\delta = 0.10\;{\rm px}$ are evidence of the higher image quality compared to the $\delta = 0.20\;{\rm px}$ of the K7M. This can be seen in the more uniform displacement field of the P1.75M and the S10M. However, the standard deviations of the S10M are larger than those of the P1.75M and the K7M. The reason for this can be observed on the one hand in the image sequences with the crack triangles, where the color-coded $\parallel {\vec t_{{\rm rel}}}\parallel$ triangles show similar values, making it difficult to determine the crack tip. On the other hand, the S10M curve shows velocities of 0 m/s. Both are caused by the evolution of the crack tip rather than image noise. If the position of the crack tip is the same in the five time steps used for velocity estimation, then the crack length is not changing and consequently there is no velocity variation. This means that, given the low spatial resolution of the S10M, the temporal resolution used is unnecessarily redundant.

Oscillations in the curve velocities shown in the three velocity plots also cause large standard deviations. The curves show an increasing and decreasing acceleration behavior of the crack tip evolution. This means that the velocity is not constant as it propagates through the specimen. Since the velocity determination includes the four temporal neighbors, an acceleration of the crack tip would introduce errors in the velocity estimation.

Comparing the precision of the velocities presented in Table 2 with the standard deviations reported in [9], no explicit increase in velocity precision was obtained, especially for the S10M. However, a more detailed evolution of the crack tip through the specimen surface could be captured. Furthermore, the average velocities obtained from the 3PB tests and from the tension tests are in a similar range (from 600 to 900 m/s). From these results it can be concluded that the standard deviations caused by the accelerations can be improved by increasing the frame rate. However, if the spatial resolution of the camera is insufficient, a high frame rate would affect the curve development. Observing the curve velocity results, the P1.75M seems to offer the best compromise between spatial and temporal resolution for tracking the crack tip. The limitation is that the field of view captured is not practical because the crack may not be propagating in the observed region. For the K7M, a longer evaluation time can be obtained by delaying the camera trigger. In the case of the S10M, better use of the camera’s field of view may slightly increase the resolution of the cracks, but the results of other experiments did not significantly improve the definition of the crack tip.

5. DISCUSSION

The applicability potential of the three cameras Phantom TMX-7510 [1.75 Mfps (P1.75M)], Shimadzu HPV-X2 [10 Mfps (S10M)], and Kirana7M [7 Mfps (K7M)] in high impact experiments is evaluated in dedicated accuracy tests. Sufficient spatio-temporal information is required to evaluate highly loaded concrete specimens before reaching tensile strength using photogrammetric methods. For this purpose, optical data between the impact and main crack initiation will be used to evaluate the material properties without significant influence of crack propagation. Furthermore, the data should improve the analysis of crack propagation using the method proposed in [9]. To this end, the information obtained from the three monocular camera systems is used on the one hand to observe the gradual deformation of the mesh surface points prior to crack formation, referred to in this work as surface particle propagation, and on the other hand to increase the precision of the crack propagation velocities.

Compared to other high-frequency measuring techniques, cameras provide a high temporal and spatial resolution for a large number of points on the specimen surface. In addition, these optical devices do not need to be attached to the experimental setup, thus avoiding vibrations caused by setup accelerations. To evaluate the quality of the application results captured with the three cameras used here, the sensor performance is evaluated using recorded image sequences. Here are the conclusions about the imaging performance:

To quantitatively characterize the expected specimen behavior of the mentioned deformation processes, the experimental setups of the three-point bending (3PB) of cuboids ($160 \times 40 \times 40\;{{\rm mm}^3}$) and SHPB of cylinders ($l = 80\;{\rm mm}$ and $\emptyset = 50\;{\rm mm}$) were selected. Appropriate testing configurations, triggering conditions, and consideration of the impact characteristics of the experimental setups were used to properly evaluate the specimen deformation. A summary analysis of the particle propagation and main crack velocity results is given below. It should be noted that although the tests were carried out under comparable conditions, the three tests are different, and the results are not directly comparable. Here are the seven main points:

It can be concluded that to capture the specimen response in the presented short-time experiments, a suitable image ratio and a convenient spatio-temporal resolution have a higher influence on the deformation results than the image quality. However, the image quality has an obvious negative influence on the image correlation process, which must be considered. Therefore, an analysis of the sensor operation can be used to understand the source of the quality errors. These errors can be reduced by using appropriate test configurations. They can also be compensated by quantifying the effect of the brightness response to increase the image SNR or to correct the reconstructed surface. However, the lack of data cannot be compensated.

6. CONCLUSIONS AND OUTLOOK

Ultrahigh-speed cameras allow the investigation of the behavior of concrete specimens tested at high strain rates, where the testing conditions to characterize the material properties depend on the specimen response. To achieve high acquisition speeds, different camera sensor characteristics are offered on the market. The BSI CMOS technology of the Phantom TMX 7510 provides a superior image quality, but the increase in the frame rate without affecting the image size remains a limitation. ISIS technology allows high temporal resolution with full dimensional resolution. Thus, the Kirana 7M CMOS sensor with a CCD memory offers an image size of $924 \times 768\;\rm px$ up to 7 Mfps and the Shimadzu HPV-X2 CMOS FTCCD sensor offers an image size of $400 \times 250\;\rm px$ up to 5 Mfps. Measurement accuracy values based on the pixel sizes of 1/50 px, 1/10 px, and 1/11 px, respectively, for each of the cameras mentioned above, were obtained using plane similarity transformations. Due to the large spatial resolution of the Kirana 7M, a higher level of performance was obtained to observe small relative step deformations before and after main crack initiation. However, due to the low sensor sensitivity of this camera, flash lamps are required. Finally, the image acquisition challenge of recording capacity is overcome by using consistent trigger times.

The generated data can be used in future investigations to further improve the estimation of the crack propagation velocity, which is additionally influenced by accelerations of the crack tip evolution due to the gained spatio-temporal resolution. The high spatio-temporal redundancy of the mesh surface point displacements, without the influence of unpredictable displacements caused by the direction of the main propagating cracks, will increase the reliability of the material analysis and allow precision statements. Therefore, using one-dimensional wave theory in solids, material properties such as the wave velocity that causes a given deformation as it crosses the tested specimen could be estimated. In addition, the stress at a given particle surface position could be derived from the estimated wave velocity, the mass density, and the particle surface velocity.