Control concepts for image-based structure tracking with ultrafast electron beam X-ray tomography

In this paper, a novel approach for tracking moving structures in multiphase flows over larger axial ranges is presented, which at the same time allows imaging the tracked structures and their environment. For this purpose, ultrafast electron beam X-ray computed tomography (UFXCT) is being extended by an image-based position control. Application is scanning and tracking of, for example, bubbles, particles, waves and other features of multiphase flows within vessels and pipes. Therefore, the scanner has to be automatically traversed with the moving structure basing on real-time scanning, image reconstruction and image data processing. In this paper, requirements and different strategies for reliable object tracking in dual image plane imaging mode are discussed. Promising tracking strategies have been numerically implemented and evaluated.


Introduction
Fast imaging techniques are essential experimental tools in fluid dynamics research, particularly for the investigation of multiphase flows. Such flows are to be found in, for example, chemical reactors, thermal power plants, oil and gas processing, refrigeration systems, fluid transport systems and others. Multiphase flows are inherently complex due to the coexistence of multiple phases with different physical properties and highly deformable interfaces; for example, in gas-liquid and liquid-liquid flows. The modelling and numerical simulation of multiphase flows with computational fluid dynamics (Brennen, 2005: 20 ff.) require validation data from experiments. For such measurement techniques with high temporal and spatial resolution are needed. Particularly, non-invasive imaging is of great attractiveness (Reinecke et al., 1998). However, today there is no multiphase flow imaging technique available, which gives a full three-dimensional (3D) view of opaque multiphase flows at high speed and resolution. As an example, we address a simple bubbly flow in a pipe or vessel, that is, a coexistent flow of continuous liquid and disperse bubble in a flow channel. As such, it would be ideal to ''see'' all bubbles in a given volume, that is, pipe or column section, at a spatial and temporal resolution that discloses the dynamics of the bubbles themselves, the dynamics of their interfaces and their geometric relations to other bubbles. Of course, it would also be of interest to know the local continuous liquid velocities. As this is all not possible today, an alternative would be to have a fast cross-sectional, that is, two-dimensional (2D) imaging, with a capability to track a moving structure, like a bubble. This would at least allow to study a single bubble and its environment dynamically. In the following, we will describe our approach to solve this task by tracking structures in a flow with an ultrafast X-ray tomography device. Before that, we will briefly review dynamic imaging techniques for multiphase flows, to provide a good start into the subject.
Today, there are a few fast flow measurement techniques available for scientists and engineers. Most prominent are high-speed cameras. However, they are seldom applicable for multiphase flows as such are inherently opaque. The same holds for ultrasound-based imaging techniques. Techniques, which are in principle suitable for multiphase flows, are electrical tomography, radiation-based imaging techniques and magnetic resonance imaging. Hence, we will first briefly review these techniques with a focus on their current capability to deliver a full picture of, for example, a bubble flow in a pipe or column.
Electrical tomography, with its variants electrical capacitance tomography (ECT), electrical resistance tomography (ERT) and electrical impedance tomography (EIT), is able to recover structures with different electrical properties in 1 Chair of Imaging Techniques in Energy and Process Engineering, Technische Universität Dresden, Germany 2 Helmholtz-Zentrum Dresden -Rossendorf, Institute of Fluid Dynamics, Germany multiphase flows. The techniques can be made quite fast and have a certain 3D imaging capability (Xie et al., 1995). However, their spatial resolution is very limited, which hampers tracking of dedicated structures.
Transmission imaging techniques based on radiation can be classified into radiographic and tomographic imaging. Fast radiography is known for X-rays (Boden et al., 2014) and neutrons (Kaestner et al., 2011a(Kaestner et al., , 2011b. X-ray radiography may be used and has been implemented in the form of Xray stereoscopic particle tracking (see below). Conventional X-ray tomography, that is being used in medicine and for non-destructive testing, is generally too slow as it requires mechanical movement of system components (Kak and Slaney, 1988). An exception is ultrafast X-ray tomography, which we address in this study (Fischer et al., 2008).
Magnetic resonance imaging (MRI) provides high spatial resolution and 3D imaging capacity. However, the fastest MRI scanners achieve up to 50 fps (frames per second) (Mu¨ller et al., 2007;Tayler et al., 2012b), which is still quite slow compared with the time scales of multiphase flows. Recent applications of ultrashort echo time MRI have attempted to track moving structures in fluidized beds (Fabich et al., 2017) and bubbly flows (Tayler et al., 2012a) at 40 fps. A main advantage of MRI is the simultaneous measurement of liquid-phase velocities that cannot be achieved by radiation-based tomography methods. However, the relatively small volume of investigation limits the application to snapshots of bubble swarms rather than visualizing their paths and dynamics over a larger range.
Besides imaging techniques, there exist also pure structure tracking techniques for multiphase flow. Such are CARPT (Degaleesan et al., 2001), PEPT (Cole et al., 2010), and X-ray stereoscopy (Nadeem and Heindel, 2018). In PEPT and CARPT, radioactively marked tracer particles are tracked over a longer period. As they are pure tracking techniques the surroundings of the tracked structures (particles) are not disclosed. X-ray stereoscopy obtains the position of highcontrast particles from triangulation. To some degree surrounding structures are disclosed in the associated X-ray images. However, as these images provide no good depth information, the 3D imaging capacity is rather low (Nadeem and Heindel, 2018).
Ultrafast X-ray tomography is a relatively new fast imaging technique that comes very close to the given requirements. It bases on X-rays, provides high frame rate of up to 8000 fps, 1 mm spatial resolution and dual plane imaging (Bieberle et al., 2012). As this technique yet has only 2D imaging capability, though in two planes simultaneously, structure tracking out of the imaging planes is not possible. To qualify this technique for structure tracking, a fast positioning system for the scanner has to be implemented and a fast data analysis strategy for controlling the system has to be developed. The latter development has been started already in our group with the transfer of the data processing onto graphic processing units (GPU) (Bieberle et al., 2017;Frust et al., 2017). This now gives short enough latency times for realtime axial scanner position control.
In the scope of the control concept the two imaging planes of the X-ray scanner act as a pair of binary structure sensors. Many control schemes for binary sensors or binary sensor networks (BSN) are known from the literature. However, their application is limited to either unidirectional continuous processes, for example, liquid level control (Wang et al., 2003), or multi-dimensional positioning problems for zerodimensional objects, for example, tracking of persons across various rooms, (Bai et al., 2015). Thus, Liu et al. (2013) proposed a strategy for extraction of velocity and size of a circular object moving at constant speed using a set of transmitter-receiver style binary sensors. However, the extracted information is used neither to control the process nor to track it actively by repositioning the sensors.
There are also various strategies for image-based control, for example, based on high-speed camera footage (Hutchinson et al., 1996). They are, however, only applicable for controlling processes within the imaging region. Combining the available concepts, we propose a control strategy for image-based control of processes occurring normal to the imaging plane(s). Here, instead of a BSN, only a pair of controllably traversable binary sensors is used.
To track a moving structure, an effective position control of the ultrafast X-ray CT scanner must be provided, employing a low-latency trajectory generation. For that a so-called model predictive control approach is most suited in which the target's position changes are suitably modelled. Furthermore, constraints regarding the maximal applicable velocity and acceleration need to be included. This requires iterative computation techniques that are critical in terms of real-time operation (Kim et al., 2007;Neunert et al., 2016). Current analytic approaches do not include such a movement model (Haschke et al., 2008;Ruppel et al., 2011). However, for tracking of a temporarily non-visible moving target, such as a bubble moving between imaging planes, this is indispensable. Thus, a time-optimal, analytic trajectory generation is proposed that complies with maximum velocity and acceleration constraints. A double-setpoint controller is employed based on the periodic modelling of the structure's movement with a parabolic position profile. Similar to the receding horizon approach, the model parameters are updated in every timestep and only the optimal control output for the next time step is used. Different from classic receding horizon control, the analytic solution allows modelling the movement up to the current control target instead of exploring only a fixed prediction horizon.

Materials
The ultrafast electron beam X-ray CT scanner As shown in Figure 1a, the ultrafast electron beam X-ray CT scanners use an electron beam that is focussed and deflected onto a circular tungsten target to generate a rapidly rotating X-ray focal spot along two staggered paths without any mechanical movement of components. Data acquisition is performed by a radiation detector that comprises two distinct rings of CdTe detector pixels arranged concentric to the X-ray target. The CT scanner under consideration offers a circular imaging area with a diameter of 190 mm and two distinct imaging planes with a geometric distance Dh c = 13 mm ( Figure 1b). The alternating scanning in two planes with a small axial offset allows the determination of structure velocities by time-of-flight methods (Barthel et al., 2015). Currently, each detector ring comprises 432 seamlessly arranged pixels, whose analogue signals are digitized with a sampling frequency of f samp = 10 6 samples/s. Cross-sectional images are acquired alternately in both imaging planes (see Figure 1c) with a maximal imaging rate of up to f im = 8000 fps.
Image reconstruction is currently performed offline. That is, the digitized signals of the radiation detectors are temporarily stored in the random access memory of the detector electronics before they are being transferred to the host computer after completion of a scan (Bieberle et al., 2017). There, image reconstruction and post-processing is done at a later time.
For structure tracking, real-time image reconstruction as well as control strategies to traverse the CT scanner are needed. Data processing pipeline systems Kopmann et al., 2016) seem to be proper tools for that. For that we evaluated three different traverse control concepts, as explained in the following.

Tracking strategy
A moving structure, in this case represented by a rising gas bubble, is fully described by the following set of onedimensional parameters (Figure 2a): the current position s s of the structure ''s'', its velocity v s , its acceleration a s and its  axial length l s . Furthermore, the CT scanner ''c'' is also fully described by its current position of the upper imaging plane (called scanning position) s c , velocity v c and acceleration a c . As constraints we have a maximal absolute velocity v max c , a maximal absolute acceleration a max c and position limits s min c and s max c . Moreover, we have the constant geometric distance Dh c between both imaging planes h 0 and h 1 and the time interval Dt cycle = 2=f im for the acquisition of an image-pair. The latter serves as a master clock for parameter extraction and storage. With each newly acquired pair of images n 2 1, ::, N max ½ the system clock is updated to t n = t nÀ1 + Dt cycle .
To track the moving structure, the scanning position s c has to be adapted by controlling the scanner's movement based on these parameters. To determine this target scanning position s target c , three strategies have been considered. For this, an additional offset parameter p\1 is introduced which is used to calculate s target c based on the current structure length l s using s target c = s s + s off i for i 2 1, 2 f g with s off 1 = À p Á l s and s off 2 = À Dh c + 1 À p ð ÞÁl s ð Þas described below. A value 0\p\1 leads to a target scanning position within the tracked structure at portion p of its length l s , that is, the structure remains visible, whereas p\0 leads to deliberate overshooting by p j j Á l s , that is, temporary non-visibility of the structure. The default values p for each of the considered tracking strategies were determined empirically (see Table 1) and may be adapted within the respective given ranges (see below) for each of the following strategies.
Single image plane tracking (SIPT, 0\p\1): In the first strategy, the CT scanner's position is controlled such that the structure remains visible in the lower image plane h 0 . Therefore, the parameters s s , l s and v s are initially determined while the structure passes both imaging planes first. Afterwards, the CT scanner drives to its target scanning position s target c at a defined portion of the structure's length l s within the image plane h 0 with a final velocity of v c = v s , as shown in Figure 2b (SIPT). The target position is therefore defined to be s target c = s s À p Á l s . In case the velocity of the structure v s does not change over the section to be investigated, the scanner has not to be accelerated anymore. Anyway, in the more likely case, where v s 6 ¼ const:, the velocity of the scanner v c has to be iteratively adapted correspondingly as described in chapter 2.3. With the structure visible most of the time in one imaging plane, this strategy is expected to provide detailed trajectory information. However, size and shape changes of the structure cannot be detected.
Dual image plane tracking (DIPT, 0\p\1): As in SIPT, the structure is initially completely scanned in both imaging planes to determine its current size and velocity. Then the CT scanner moves upward, thereby overtaking the structure with the upper imaging plane and partially with the lower imaging plane until the lower image plane is a distance p Á l s above the bottom of the structure (Figure 2b-DIPT). There, the CT scanner accelerates with Àa max c such that the upper image plane is positioned at distance p Á l s below the top of the rising structure. For very large or slowly rising structures, this leads to intentional downwards movement of the scanner. Thereby, the number of sampling points t k is maximized (see section 2.3). This procedure is then repeated. With DIPT the structure is always in the axial region between the upper and lower image plane and visible for almost all the time. By letting the structure frequently pass through both imaging planes, its shape and size can be continuously measured. However, this strategy is still somewhat sensitive to fast changes of the structure's geometry that may result from, for example, flow disturbances.
Dual image plane re-tracking (DIPRT, p\0): In the third strategy, the structure is repeatedly scanned completely by the CT scanner in both image planes. As described for the DIPT strategy, the target scanning position again alternates between the two target offsets s off 1 and s off 2 but this time with the offset parameter p\0. Thereby, the scanner drives above the structure and scans again the structure's parameters. Afterwards, the scanner accelerates with Àa max c to let the structure pass completely through both imaging planes once more and repeats the cycle. Employing this strategy, the structure's shape and size changes can be characterised. However, information about its trajectory is limited to the intermediate scans. Further, the structure leaves the scanning planes multiple times and needs to be rediscovered.

Parameter extraction of the moving structure
No matter which tracking strategy will be applied, parameters of the investigated object have to be extracted initially from the image pair data. However, as recognizable from Figure  1c, in a single image plane, a small and slowly rising structure leads to the same image sequence as a tall and fast rising Table 1. Parameter space for the numeric simulations of the behaviour of the complete system, that is, CT scanner and control unit, as depicted in Figure 9.
structure. Thus, for the extraction of the structure velocity, the images from the second imaging plane h 1 need to be employed to track the boundaries of the structure whilst passing through both imaging planes (see Figure 3a). The time interval Dt 1 = t 2 À t 1 (see Figure 3a) is initially determined as the time between the structure's front entering the lower plane h 0 and upper plane h 1 . Afterwards, the velocity of the structure's boundary can be calculated by Secondly, the time interval Dt 2 = t 3 À t 1 between the entering of the structure's front at the lower image plane h 0 and the structure's back at the lower image plane h 1 is determined. As can be seen in Figure 3b, the current axial length of the structure l s can then finally be calculated by using the latest previously estimated velocity v s . Similarly, the structure's lower boundary can also be used to determine these parameters.
By following an axially moving structure, its boundaries are tracked every time one crosses one of both image planes. Between these crossing events at times t k , the structure's length is assumed to be constant. At each such event, an updated structure velocity v s value can be calculated and, thus, the structure's acceleration a s = v s t k ð Þ À v s t kÀ1 ð Þ ½ = t k À t kÀ1 ð Þ . Consequently, continuous updates of the estimated structure parameters s s , v s , a s and l s are realized by evaluating the crossing events together with both their corresponding time-stamp t k and their respective CT scanner position s c (see Figure 7).
To distinguish between the structure's upper and lower boundary, six different CT scanner states S have been defined, as shown in Figure 4.
States 1, 2 b and 3 can be classified solely from visibility information in the reconstructed cross-sections and are therefore defined as certain states. To differentiate between states 0, 2 a and 4, the state history of the CT scanner must be considered. These states are therefore classified as uncertain states. Possible transitions between states are shown in Figure 5.
Each transition is used to identify the tracked structure's upper or lower boundary, respectively. For example,  Horizontal black lines represent the imaging planes of the CT scanner. Note, that states 2 a and 2 b are only valid for structures smaller than or greater than the distance between the imaging planes, respectively. transition 0 ! 1 indicates the structures upper boundary and 3 ! 4 the structures lower boundary at the time of scanning. However, because uncertain states can only be determined based on the latest estimated structure properties, like v s and l s , false state estimations may occur due to disturbances like turbulence or vibrations. This may lead to invalid transitions as shown in Figure 6 (dotted line).
In case such an invalid transition is detected, the system falls back to the latest certain state at time t m and retroactively corrects all state estimates and structure property extractions up to the current time-stamp t n . Thus, the CT scanner states must additionally be acquired and saved at each state transition with their respective time-stamp. Each of these events is considered as a sample point t k for the parameter extraction. Based on the continuously updated structure parameters the CT scanner's acceleration profile can be properly adapted (see Figure 7).

Motion planning for the CT scanner
Depending on the above given tracking strategy and the latest parameter estimation of the structure, a new target position s target c is calculated for each clock cycle (see Figure 2). To reach the target height in the minimum positioning time interval Dt trav , acceleration and deceleration phases have to be performed with the systems maximal applicable acceleration  The transition from state 4 at t n is invalid (dotted line). Therefore, the estimated system states are retroactively corrected from the last certain state and time stamp (S3 at t m ) such that a valid sequence of state transitions is ensured (dashed line). Feature estimations from false state transitions are also corrected.
value a max c . Thus, the generated trajectory for the CT scanner's movement needs to comply with these boundary conditions The structure's movement profile is described by For better readability and without loss of generality, terms in the form s s t 0 ð Þ are shortened to s s and the initial time t 0 is set to 0 in the following. Based on the structure's movement profile, a similar profile for the target position s target c of the CT scanner can be calculated using s target c t ð Þ = s s t ð Þ + s off i with i 2 1, 2 f g as explained in section 2.2. Because the three discrete possible acceleration values a c for the CT scanner are predefined (see equation (6)), only the time intervals for acceleration and deceleration phase need to be calculated to provide the time-optimal trajectory. This optimal acceleration time Dt acc is calculated explicitly by solving the given parabolic position and linear velocity profiles (see Figure 8) Therein, the directional term dir 2 À1, 1 f g is used to switch the order of acceleration and deceleration depending on the current situation at each master clock event. Using the wrong dir-value leads to a transition time Dt acc \0. If this is the case, the calculation for Dt acc is repeated with dir = À dir. If the calculated acceleration interval Dt acc does not lead to a violation of the predefined maximum velocity v max c , the corresponding deceleration time can be calculated. Otherwise, an intermediate phase segment with constant and maximum velocity of duration Dt lin must be added between acceleration and deceleration interval ( Figure 8b). Therefore, the acceleration time interval Dt acc must accordingly be updated at first by considering the current velocity v c of the CT scanner. Secondly, the corresponding deceleration interval is calculated by Lastly, the intermediate time interval Dt lin is calculated from the given movement profiles by Figure 7. Concept of evaluating structure's front and back using image-pairs and the history of the CT scanner states using the example of the initial investigation of structure parameters required at the beginning of each of the aforementioned structure tracking strategies.
Finally, equation (15) on which the scanner will hit its target scanning position. The trajectory is updated every master clock cycle and passed as input to the scanner's positioning unit. After presenting different tracking strategies and the parameter extraction procedure for the observation of dynamic structures, the entire control concept (Figure 9) can be concluded. As input parameters the continuously reconstructed image-pair stream as well as the current position of its imaging planes s c are required.
Each time the CT scanner state S is changed, the structure's parameters are updated. Based on these extracted parameters as well as the limits of the positioning system, the CT  scanner's motion is planned using one of the three introduced tracking strategies.

Results and discussion
To investigate the feasibility of the proposed control strategies the behaviour of the entire system, that is, CT scanner and control unit as depicted in Figure 9, has been simulated for various flow scenarios. For this, an artificial 1D logical phantom vector which resembles predefined structure motion and length profiles (see Figure 10) was used.
To cover a wide range of technical applications different movement and structure length profiles have been simulated (see Figure 10). This set of parameters includes linear, accelerated, decelerated and up/down movement as well as constant, linearly increasing / decreasing and fluctuating structure lengths. The used values for the study are compiled in Table 1 with their respective indices. Combining all parameter variations, a total number of 4536 different parameter sets have been simulated.
During the simulation, the path and the velocity of the structure phantom is calculated for each current time step t n considering also the input delay Dt in , which accounts for data transfer and data processing delays. According to Frust (2016), the image reconstruction latency is expected to be 4:3 ms at maximum for the default imaging rate of f im = 2000fps. Data transfer times are estimated to be 0:2ms based on the available bandwidth. For structural feature extraction an additional latency of 0:5ms is assumed based on the usage of a thresholdbased binarization data processing stage. The total input latency for the default image-pair rate is, therefore, estimated to be Dt in = 5:0ms (see Figure 9). A larger input delay Dt in is also considered because the recognition and rediscovery of a structure may require multiple pairs of images. Furthermore, the output delay Dt out , that is, the time interval between generating the acceleration signal and realizing the corresponding movement, is estimated to be 1ms based on commercially available motor controllers (SEW-EURODRIVE, 2010).
In the real positioning system, the position s c is determined using a wire potentiometer with a measurement uncertainty of Ds meas c = 60:1mm. This offset was added as Gaussian noise to the current system position s c at each clock cycle in the simulations. An additional larger position uncertainty Ds meas c = 60:5mm was simulated to account for mechanical  (7)) and b) length profiles (constant (1,2), linear increase (3,4), linear decrease (5,6), fluctuating (7,8)) along the test section. Dotted lines represent Dh c . Quantitative definitions are given in Table 1. offsets, for example, vibrations or sagging of the wire, in the real setup. Height information time delay is negligible. A time step discretization, that is, cycle time of Dt cycle = 1ms is used to fit the typical acquisition rate of f im = 2000fps. The CT scanner's position limits are s min c = 500 mm and s max c = 2500mm: The CT scanner starts at standstill, that is, v c t 0 ð Þ= 0, at an initial height s c t 0 ð Þ= 500mm and comes to a complete stop at its maximum height s max c . Its maximum velocity and acceleration were set to v max c = 500 mmÁs À1 and a max c = 500 mm Á s À2 , respectively. The structure's initial height is s s t 0 ð Þ= 0mm. All simulations were carried out in Matlab. All parameter sets with Ds meas c . 0 were simulated 100 times to get statistically significant information about the influence of position noise Ds meas c whilst retaining manageable simulation times (about 5 seconds / simulation).

Structure parameter extraction
To evaluate the tracking strategies regarding velocity information extraction, the momentary velocity estimations v s t k ð Þ at each sample point t k (see Figure 5) were compared with the momentary nominal velocities v nominal s , that is, the predefined velocities of the phantom vector, for each set of parameters being simulated. The median velocity estimationṽ s t k ð Þ of each simulation was selected and averaged with the estimate of all 100 simulations of the same parameter set. The median value was selected instead of the mean value because single estimations may become implausibly large or small due to short durations between sample points or numeric limitations, for example, division by length scales close to floating point precision. Results are shown in a parity plot in Figure 11.
Results at all sample points show an agreement between nominal velocity and median measured velocityṽ s t k ð Þ with deviations below 2%, which is acceptable for this application. Furthermore, velocity estimates have been identified to be insensitive to structure length l s , system delay Dt in=out and position noise Ds meas c in the tested ranges. To evaluate the length estimation quality the momentary length estimations l s t k ð Þ for each simulation were compared with the momentary nominal structure length l nominal s as explained previously for the velocity estimates. Median length estimation resultsl s t k ð Þ show good agreement with typical deviations better than 10% (see Figure 12). Length estimation deviations tend to increase for fluctuating structure lengths which results from the limited number of sample points for the length determination, that is, relatively low sample frequency compared with the fluctuation frequency. Length estimates are insensitive to system delays Dt in and Dt out as well as position uncertainty Ds meas c in the tested ranges.

Comparison of control strategies
As depicted in Figure 13a Figure  13b). This is due to false state estimates which cannot be corrected with the limited data provided by SIPT, especially for structure lengths shorter than the distance between the imaging planes Dh c because a differentiation between states 0 and 2 a is impossible (see Figure 4) Moreover, DIPRT's length estimates show a bimodal distribution (see Figure 13b). Further investigations of length estimations for DIPRT in different structure length profiles reveal that this bimodal distribution occurs only for fluctuating structure length profiles (see Figure 14a).  This bimodal distribution occurs for all simulated movement profiles, delays Dt in and Dt out as well as position noise levels Ds meas c in combination with fluctuating structure length (see Figure 14b). It is especially apparent for initial structure velocities v nominal s t 0 ð Þ= 50mm Á s À1 . The reason for this distribution is a systematic underestimation of fluctuating structures with mean length l s \Dh c combined with an overestimation of fluctuating structures with mean length l s ø Dh c . This tendency is most apparent in DIPRT due to its narrower distribution of length estimates. In SIPT and DIPT, the wider distributions do overlap and thereby conceal this tendency when averaging all test cases.
Increasing input latency Dt in leads to an overall increase in the velocity estimate's standard deviation in each control strategy from sṽ s t k ð Þ=v nominal s À Á = 0:015 for Dt in 2 0, 0:005s f g to sṽ s t k ð Þ=v nominal s À Á = 0:035 for Dt in = 0:010s. Variation of output delay Dt out and position noise Ds c do not show significant deviations (sṽ s t k ð Þ=v nominal s À Á ł 0:03) in the simulated range of values. An acceleration switching frequency of 10Hz was sufficient for all simulated parameter sets, that is, the electro-mechanical positioning system requires a corresponding cut-off frequency. The influence of different time discretization Dt cycle was only briefly investigated and did not show noteworthy differences.

Conclusion
Different tracking strategies for the application of ultrafast Xray tomography for studying moving structures in multiphase  flows have been evaluated. A control scheme was proposed that provides a defined CT scanner positioning movement based on currently acquired image-pair data. Therefore, three different tracking strategies, namely Single-Image Plane Tracking (SIPT), Dual-Image Plane Tracking (DIPT) and Dual-Image Plane Re-Tracking (DIPRT), were discussed and evaluated by numerical simulation. All strategies promise very good velocity tracking results. However, SIPT is preferred for cases with constant structure shapes and lengths, for example, tracking of a tracer particle. To extract length and shape information for time-variable structure length and shapes, DIPT is more suitable for low disturbances. The DIPRT approach is preferred for detailed shape determination at high turbulence cases. Simulations have shown that DIPT and DIPRT were more robust than SIPT concerning the influence of time delays and position uncertainty.
Future work will focus on reliable structure recognition from the image pair data. Moreover, further post-processing steps to increase state estimation quality, and therefore feature extraction quality, using computationally complex techniques like particle filtering are of high interest.

Declaration of conflicting interests
The author(s) declared no potential conflict of interests with respect to the research, authorship and/or publication of this article.