This letter presents a novel computationally efficient and robust pattern tracking method based on a time-encoded, frame-free visual data. Recent interdisciplinary developments, combining inputs from engineering and biology, have yielded a novel type of camera that encodes visual information into a continuous stream of asynchronous, temporal events. These events encode temporal contrast and intensity locally in space and time. We show that the sparse yet accurately timed information is well suited as a computational input for object tracking. In this letter, visual data processing is performed for each incoming event at the time it arrives. The method provides a continuous and iterative estimation of the geometric transformation between the model and the events representing the tracked object. It can handle isometry, similarities, and affine distortions and allows for unprecedented real-time performance at equivalent frame rates in the kilohertz range on a standard PC. Furthermore, by using the dimension of time that is currently underexploited by most artificial vision systems, the method we present is able to solve ambiguous cases of object occlusions that classical frame-based techniques handle poorly.
Real-time pattern-based tracking is a fundamental component of various applications in artificial vision. It deals with the problem of establishing correspondences between a known model of an object consisting of a set of 2D points and a target appearing on the image plane. The goal is to estimate the relative pose between the model and the target. Current pattern tracking methods are designed to work on images, or frames, acquired at a fixed rate. Image acquisition is usually limited to the order of tens of milliseconds in real-time applications. This drastically restricts the ability to track correctly high-velocity objects. Increasing the frame rate is often not a solution as the increasing amount of acquired data sets a limit to real-time computation. This is currently the bottleneck of several mobile robotics applications, where there is always a trade-off to find between the frame rate and the computational load allowed.
A recent and evolving branch of artificial vision exploits the unique characteristics of a novel family of asynchronous frame-free vision sensors whose principles of operation are based on abstractions of the functioning of biological retinas (Lichtsteiner, Posch, & Delbruck, 2008; Posch, Matolin, & Wohlgenannt, 2011; Lenero-Bardallo, Serrano-Gotarredona, & Linares-Barranco, 2011). These cameras acquire the content of scenes asynchronously. Every pixel is independent and autonomously encodes visual information in its field of view into precisely time-stamped spike events. These cameras allow the derivation of a new methodology for computation that operates on pixels or areas of the image at the time they convey relevant (dynamic) information. Processing can be performed on standard digital hardware and takes full advantage of the precise timing of the events. The increasing availability and the improving quality of these sensors open up the potential to introduce a shift in the methodology of acquiring and processing visual information in various demanding machine vision applications (Perez-Carrasco, Serrano, Acha, Serrano-Gotarredona, & Linares-Barranco, 2008; Benosman, Ieng, Clercq, Bartolozzi, & Srinivasan, 2012; Benosman, Ieng, Posch, & Rogister, 2011; Serrano-Gotarredona et al., 2009).
As we will show, this biomimetic asynchronous acquisition allows to introduce a novel computationally efficient and robust visual real-time pattern-based tracking method that relies on the accurate timing of individual pixels’ response to visual stimuli. We will further show that asynchronous acquisition allows us to develop a tracking technique that can follow patterns at an equivalent frame rate of several kilohertz, overcoming occlusions at the lowest computational cost. The high temporal precision of the presented methodology does not necessitate motion estimation filters to ensure the smoothness of the motion trajectory or the use of a priori kinematics models.
There has been a lot of research on visual pattern-based tracking. The algorithms differ in three main aspects: the appropriate representation for the patterns of interest, the transformation model used to transform the source model to match with the target pattern, and the tracking criterion used to estimate the optimal transformation parameters given a pattern representation and a transformation model. Point cloud is the most commonly used pattern representation. Iterative closest point (ICP), as the name suggests, iteratively uses a point set delineating the object contour to match an acquired data point set by searching their nearest correspondences (Besl & McKay, 1992). Basic curve or surface elements can also be used for matching. Robust point matching (RPM) is a method to deal with nonrigid shapes by treating the point matching as an optimization problem to preserve local neighborhood structures (Zheng & Doermann, 2006). The shape contexts method attaches a descriptor to each point during the matching process (Belongie, Malik, & Puzicha, 2002). Moreover, distance functions or distance transforms are applied for shape alignment in Huang, Paragios, and Metaxas (2006). A level-set-based implicit shape representation framework is presented in Cremers, Osher, and Soatto (2006) for human body tracking and object segmentation. Finally, a parametric curve or surface requires explicitly parameterizing shapes, which makes arbitrary form shape representation non-trivial (Abd El Munim & Farag, 2007).
The transformation between the data set and the model set may be separated into two categories. Global tracking is performed between the known model set and the data set to express their overall geometric relations by satisfying some tracking criterion. As an example, a modified ICP relies on affine transformations (Rusinkiewicz & Levoy, 2001). The second category deals with a local deformations issue to provide a more accurate match. Thin plate spline (TPS; Bookstein, 1989) is such a method for parameterizing nonrigid spatial mapping and modeling deformation. The combination of robust point matching and thin-plate splines forms a more general framework known as the TPS-RPM method (Yang, 2011). Other known methods are free-form deformations (FFD; Rueckert et al., 1999) and radial basis functions (RBF; Fornefett, Rohr, & Stiehl, 1999). Complete solutions usually combine a global tracking and a local deformation transform to achieve the best tracking result. Local deformations are beyond the scope of this letter and will be dealt with in future work. In this letter, we focus on introducing a novel methodology for visual computation using the notion of visual events for shape tracking.
Another key characteristic of the proposed approach is its ability to deal with highly dynamic situations in real time. Generally the most time-consuming part of tracking algorithms lies in determining the correspondences between the data and the model. K-dimensional (kd) tree allows the matching procedure in ICP to be reduced from an average complexity of to . A multiscale or coarse-to-fine scheme is illustrated in both Jost and Hugli (2003) and Kim & Kim (2010) to accelerate the process. Hardware implementations (e.g., FPGA) are often employed to overcome performance limitations toward real-time operation (Dandekar & Shekhar, 2007). However, hardware development cycles are usually longer than that of software dealing with the same algorithm complexity, and complex tracking algorithms are difficult to express in a hardware description language such as VHDL. Although many techniques improve the processing speed, most of them are still not adapted for real-time high-speed applications. The frame-based tracking’s computations remain bounded by the acquisition frequency of the camera used. Furthermore, most of the developed techniques are power and time-consuming and are mostly used offline. The visual shape tracking technique we describe offers an alternative in performing pattern tracking using the timing of changes of individual pixels’ illuminance as a computational input. We will show that this technique is naturally highly robust to occlusions and that it allows real-time processing at a very high temporal accuracy and very low computational cost.
2 Time-Encoded Imaging
Biomimetic, event-based cameras are a novel type of vision devices that, like their biological role models, are driven by events happening within the scene, and not like conventional image sensors by artificially created timing and control signals (e.g., frame clock) that have no relation whatsoever to the source of the visual information (Delbruck, Linares-Barranco, Culurciello, & Posch, 2010). The pattern tracking method that we present is designed to work on the data delivered by such a time-encoding sensor and takes full advantage of the superior characteristics—most important, the ultra-high temporal resolution and the sparse data representation.
The asynchronous time-based image sensor (ATIS) used in this work is a time-domain encoding image sensors with quarter video graphics array resolution (Posch et al., 2011; Posch, Matolin, & Wohlgenannt, 2008). The sensor contains an array of fully autonomous pixels that combine an illuminance change detector circuit and a conditional exposure measurement block.
As shown in the functional diagram of the ATIS pixel in Figure 1, the change detector individually and asynchronously initiates the measurement of an exposure/gray scale value only if—and immediately after—a brightness change of a certain magnitude has been detected in the field of view of the respective pixel. The exposure measurement circuit in each pixel individually encodes the absolute instantaneous pixel illuminance into the timing of asynchronous event pulses, more precisely into interevent intervals.
Since the ATIS is not clocked like conventional cameras, the timing of events can be conveyed with a very accurate temporal resolution on the order of microseconds. The time domain encoding of the intensity information automatically optimizes the exposure time separately for each pixel instead of imposing a fixed integration time for the entire array, resulting in an exceptionally high dynamic range and improved signal-to-noise ratio. The pixel-individual change detector–driven operation yields almost ideal temporal redundancy suppression, resulting in a maximally sparse encoding of the image data.
Figure 2 shows the general principle of asynchronous imaging spaces. Frames are absent from this acquisition process. They can, however, be reconstructed, when needed, at frequencies limited only by the temporal resolution of the pixel circuits (up to hundreds of kiloframes per second). Static objects and background information, if required, can be recorded as a snapshot at the start of an acquisition, while moving objects in the visual scene describe a spatiotemporal surface at very high temporal resolution.
3 Event-Based Pattern Tracking Algorithm
The event-based pattern tracking updates iteratively the model location and orientation to match as close as possible to the target the 2D image plane based on the arrival of events. The algorithm is tailored to track objects even if they are subject to affine transformations. In a first stage, the novel algorithm making full use of the sparse data will be shown. In a second stage, important spatiotemporal properties are derived from the generated events and will be shown how to use them for multiple object occlusions disambiguation.
3.1 Spatial Matching
3.2 Event-Based Estimation of Geometric Transformation
Once the matching stage has been performed, it becomes possible to update the geometric transformation between the events and the model. is defined as the spatial location of the center of gravity of the model (see Figure 3). In the calculation afterward, the transformation is performed with respect to the center of gravity; thus, we define and .
where is the rotation in the z-axis, about an angle of . The superscript T in equation 3.7 stands for the transpose operation.
sx and sy are scaling factors—in other words, the change of the size of the object—that applied to the two axes with the following definitions:
To eliminate the orientation and retrieve only the scale information, we set to represent the size of the object at a particular moment and as the size of the model along the x- and y-axes.
where is the scale adjusting parameter and i is the iteration index.
It is important to emphasize that the spatiotemporal approach performs an iterative computation for each incoming event rather than recreating images by accumulated events. This event-based approach significantly reduces the computational load as it is not processing previously processed data. It allows computations at temporal resolutions equivalent to frequencies of several kHz, as we show in section 4.
3.3 Occlusions in Multiple Object Tracking
When multiple objects cross and occlude each other, both spatial and temporal constraints can be used, separately or in combination, to disambiguate these overlaps. An ambiguous event is an event that can possibly be attributed to several objects. In actual implementation, the distances between the event and the closest matching point on each interested object are calculated. If only one of the distances is below a threshold (e.g., 3 pixels), then this event is attributed to that object; otherwise it is considered ambiguous.
3.3.1 Spatial Constraints
Considering spatial information, several strategies can potentially be used to assign incoming events to multiple overlapping shapes:
Assign to closest. An incoming event is assigned to the closest model. In theory, this should work correctly if the incoming event perfectly matches the model shape and if the data contain no or little noise.
Reject All. This approach simply throws away all ambiguous events. This appears to be a logical solution since a priori, no clear decision can be made. However, the tracking may fail as the true dynamics of shapes is lost.
Update all. All the models around an ambiguous event are updated,
Weighted update. An ambiguous event contributes to each model according to a weighted update proportional to its distance to each shape.
3.3.2 Temporal Constraints
The high temporal resolution of event-based acquisition provides crucial additional information to solve ambigous situations. A shape’s event rate contains information about the tracked object and is partly encoding the object’s dynamics.
An object indexed by k, moving at a constant speed within a short period, is characterized by an event rate rk. An event is tested for its temporal consistency with each object k. If is the time of the last event assigned to the object k, the duration should be close to the rate rk if the belongs to the same object k. In short, the event is assigned to the object that leads to the Ck closest to 1.
3.3.3 Spatiotemporal Constraints
In case of two moving objects that overlap, the set of ambiguous events in the spatiotemporal space is a line. Compared to the large set of events describing the moving objects, occlusions concern only a tiny subset of them, thus providing more robustness to the tracking process.
The complete event-based tracking algorithm is summarized in algorithm 1.
In the first stage, experiments are carried out in a controlled laboratory environment to prove the method and study its properties. In the second stage, outdoor multitracking experiments are done to investigate different spatial and temporal constraints for disambiguating occlusions.
4.1 Controlled Environment Experiments
A tracking example on event-based data of a moving star shape is shown in Figure 5. The camera acquires data of the shape fixed on a rotating disc. The camera faces the disc such that the geometric relation between the model and the acquired events is roughly isometric. The disc rotates at a speed of 670 rpm. The model (similar in the following sections) is generated manually by selecting pixels in a frame. It contains points—6 on each edge. A distance threshold of 6 pixels is applied to each incoming event for the matching stage in order to eliminate the impact of noise and reduce the computational load at this stage. As shown in Figure 5, the algorithm succeeds in tracking the rotating shape efficiently despite of the high rotation velocity.
In order to provide a comparison with conventional frame-based acquisition and processing, two classic frame-based shape tracking techniques have been tested on artificially created frames. The first relies on a classical approach using singular value decomposition (SVD; Challis, 1995), and the second is the ICP algorithm.
The SVD method requires sufficient measurements to produce correct estimations. We then generate frames from a event accumulation period. The is determined from experiments.
The frame-based ICP uses images generated by accumulating events during periods of . The process simulates a hypothetical high-speed frame-based camera running at . The generated image implies that points can only be processed in batch mode on the whole frame. In order to ensure a fair comparison, only pixels’ locations remaining after frame differencing are considered. The has been determined experimentally so that the generated frame contains sufficient events in order to have the full shape of the tracked target.
To evaluate the precision of event-based tracking, the points’ fitting accuracy is measured by computing the mean distance between the model set and the location of active events every . Figure 6 shows the tracking error with respect to time. The event-based tracking with iterative isometry estimation, closed-form rigid transformation based on SVD, and a purely frame-based ICP achieves mean errors of , , and pixels with standard deviations of , , and pixels. Higher temporal resolution leads to more precise tracking. The differences between successive tracking results are less significant when the temporal resolution is finer (Ni, Pacoret, Benosman, Ieng, & Régnier, 2011). A periodic pattern can be observed on the error curve, which originates from the repetitive rotations of the stimulus. The tracking error demonstrates a good reproducibility and reliability of the algorithm. This error is due to the pixels’ square array geometry combined with the limited spatial resolution of the sensor that does not provide an isotropic response with respect to the angular position of the shape.
Figure 7 shows the trajectory of the translation and rotation estimates of the rotating star. The solid curves show the translation with respect to the x- and y-axes in the image plane. The sinusoidal curves indicate a repetitive circular motion in the 2D plane. The dotted curve records the rotation (modulo ), and the trajectory corresponds logically to a constant angular velocity. In order to provide a measurement of our tracking algorithm, the ground truths are analytically calculated by using the fully calibrated speed of the DC motor that can be compared to the measured tracking data. The errors are shown in Figure 8. The solid lines are errors of translations, and the dotted ones are those of rotation. The mean spatial and angular errors and their standard deviations are, respectively,
One can see that the spatial errors are subpixelic.
4.2 Impact of Number of Points
This section focuses on the impact of the size of the model on computational costs and precision.
Figure 9 shows the link between the size of the model and the maximum frequency that can be achieved. The maximum equivalent frame rate is defined as the inverse of the computational time for processing one event. In our application, a size of model points can provide a detection frequency of around equivalent frame rate. Up to 2000 model points, the pose can be updated at an equivalent frame rate of . The experiment shows that for a size of 60 to 70 model points, the algorithm is able to track in real time a shape moving at a speed up to 1250 rpm. As a guideline for selecting the size of the model, the corners on the object contour should always be used. On a straight edge, however, the number of points can be reduced, as this will usually not have an influence on the final tracking precision.
The mean tracking error with respect to the size of the model is also illustrated in Figure 9. When the point number increases, the error converges to 0.84 pixel due to the spatial resolution limit of the event-based camera. Logically, larger models achieve better tracking precision but require more computational effort. A size of 60 to 100 points is a good compromise for maintaining a reasonable precision (around 0.90 pixel) and preserving a high tracking frequency (around ). The program is running on a computer equipped with Intel Core i5 CPU at 2.8 GHz, taking 25% the CPU occupation.
4.3 Affine Tracking
To examine the algorithm’s affine tracking capability, three objects are used as the target objects: a rotating H, a star, and a car (see Figure 10). During the experiment, the distance and the perceptive angle between the object and the event-based camera are continuously changing. The objects are de-formed and resized as shown in Figure 10. In Figure 11, the scaling factor with respect to the original size during the car tracking is shown. Different scaling ratios along two axes can be observed. This difference is due to the fact that the changes of perspective are more often performed along the vertical axis during the test. This experiment shows that the algorithm can efficiently track complex planar objects moving in 3D space with a mean tracking error of pixels and a standard deviation of pixels (see Figure 11). Due to the flexibility added by the scaling factors, the affine estimation appears to have more accurate point matching between the model and the active events than a rigid transform, an expected result.
4.4 Object Tracking in Outdoor Scenes
In this experiment, real-world car traffic data acquired with the asynchronous camera are used. As shown in Figure 12, vehicles are moving simultaneously on different lanes. In order to provide a fair comparison, the same sequence has been used for both the time-oriented algorithm and the classic ICP frame-based method as appearing in Du, Zheng, Ying, and Liu (2010). The gradient adjusting parameters , , and as described in section 3 are experimentally set to , , and , respectively. While the event-based method uses the timing of each incoming pixel’s value at a precision of 1 s for computation, the classic frame-based ICP relies on the gray-level images reconstructed at a frame rate of fps. Each image is preprocessed. Moving edges are obtained by computing the frame difference between successive images followed by a thresholding operation. Frame-based ICP is then applied for object tracking. No postprocessing motion filters are supplemented to smooth the motion trajectory.
Figure 12 shows the results of the two algorithms applied to follow multiple targets—in this case, two objects simultaneously. A van and a car have been selected as tracking targets. The car model has 67 points, and the van model has 102 points. Templates are generated by manually clicking points on an acquired image. Column a of Figure 12 shows the event accumulation frame generated by each vehicle superimposed with the matched shape using the event-based algorithm. The same results are superimposed on the corresponding gray-level images in column b. Finally, column c shows the tracking results of the frame-based ICP. The tracking errors for each method are given in Figure 13. The image patches correspond to the critical situations encountered by both techniques. Each image patch corresponds to several local maxima of the error curve. For better readability, we chose to label tracking errors only from the van experiment. Results from other moving objects exhibit the same range of errors. Due to its size and path in the scene, the van encounters several severe occlusions.
Large static and mobile occlusions appear frequently in this sequence. Images shown in Figure 12 are selected for highlighting the difficult situations. The van is entering an occlusion zone due to a tree covering part of the lane. The situation can be seen more clearly in a1, a2, and b1 in Figure 13. The third row shows a case of dynamic occlusion due to several moving cars surrounding the “van”. The situation can also be seen in frames a4, b2, and b3 in Figure 13. Finally, the last row shows the van in a combined case of static and dynamic occlusions caused by a light pole and multiple cars that are moving simultaneously, also observable in frames a2, a5, b3, and b5 in Figure 13.
The results in Figure 12 and their corresponding errors in Figure 13 show that the computation using precise timing is more stable than the frame-based one, which loses the shape’s orientation several times. The mean tracking errors for event- and frame-based methods are around 0.86 and 5.70 pixels, respectively, with standard deviations of 0.19 and 2.11 pixels. The frame-based ICP method is not always capable of providing a correct match. Wrong point correspondences lead the minimization process to local minima. The method suffers more from static or mobile occlusions (tree, lamp posts, automobiles, and so on) with errors raising up to more than 10 pixels. A high frame rate acquisition (100 Hz in this case) is not always sufficient to overcome occlusions. The dynamic content of the asynchronous, event-based sensor data provides more stable input. Static obstacles do not generate events; therefore, they have little impact on the tracking process.
4.5 Resolving Occlusions
In this section, we compare different ambiguity-resolving strategies for multiple object tracking. The importance of the concept of dynamics brought by the event-based processing framework and the importance of timing will be studied. The spatial and temporal constraints presented in section 3.3 are examined with the same traffic sequence as above.
A van and a car (as shown in Figure 15) moving in the same direction but with different speeds are tracked simultaneously. During a specified period, the van and the car are superimposed; then the van overtakes the car. Objects other than the van or car are considered as background noise. (Its processing method will be shown in section 4.6.)
In Figure 14a, the different distances between the center of gravity of the two vehicles are shown. Ambiguities arise when these curves have similar values, meaning that both vehicles are spatially overlapping (at time 2.2 s to 2.9 s). It is clear that in that case, the use of spatial information is not sufficient unless the size of the common region is very small. Figure 14b shows the mean event rates of each vehicle. The results prove that both vehicles remain separable during all the sequence, based on event rates of around 3000 events per second for the car and 5000 events per second for the van. Clearly, time and space provide complementary information, and the temporal domain data constitutes a valuable supplement to classic spatial-only methods.
Figures 15a to 15f plot the estimated model speeds (solid lines) of the van (blue) and the car (red) and the ground truth speeds (dashed lines). The ground truth is obtained manually by identifying the corresponding points on successive images and then estimating the speed according to the time differences. The mean speeds for the two automobiles are 42.0 and 24.3 pixels per second, respectively. The image patches in Figure 15 are selected to present the most interesting moment to study the differences between each method. The beginning of each curve provides similar results because no occlusion has happened yet. Figure 15, frame a1, shows the beginning when the detection results are equally good for all.
The experimental results of different strategies are analyzed separately:
Assign to closest. No special processing is performed in this case. In Figure 15, frame a2, events that should belong to the van mistakenly match the car model. It can be observed that the car model moves toward the van’s location, and the van model is consequently lost. Different sources of noise may contribute to this failure, with the most important being the model construction noise as the model shape does not exactly match the event contour and the contour varies during time. In Figure 15, frame a3, the car finally returns to the correct position but the van is lost from tracking.
Reject all. Throws away all the ambiguous events. This method is shown to be unsuccessful. In Figure 15, frame b1, the van starts to overtake the car. In frame b2, the car is lost from its original location and then, in frame b3, is attracted by another black car with the same shape right behind. During this process, the estimated speed of the van is close to the ground truth, although oscillations appear. The speed of the car, however, falls far below the ground truth, because when all the ambiguous events are rejected, the dynamics introduced by the events are not used for computation. Since the van is much larger than the car, only a small fraction of its events is rejected. It can then still be tracked efficiently.
Update all. Updating all the possible models around the ambiguous event shows better results. Both vehicles are tracked and no shape is lost (see Figure 15, frames c1 and c2). At the end of the sequence, the events generated by the van attract the car model, so that the resulting car trajectory deviates slightly toward the van’s position and the estimated van speed becomes lower as the car speed becomes higher (see Figure 15, frame c3).
Weighted update. The dynamics introduced by the ambiguous events are distributed to all the colliding objects with weights adjusted according to the distances. The van and the car are tracked without loss, and the speed curves better match the ground truth (see Figure 15, frames d1 and d2). However, differences between the estimated car speed and the ground truth are still present at the end of the sequence (see Figure 15, frame d3). The reason is the same as in the ``update all'' strategy. However, the error is slightly lower than the ``update all'' strategy. This is the preferred strategy among the spatial solutions when multiple objects cross or collide.
Temporal constraint. Using time as the only element to segment objects according to equation 3.13 is a strategy that can give acceptable results only in some constrained cases. As expected, it cannot produce reliable results when targets have similar event rates. Even in the cases where the targets have different events rates, the temporal constraint alone should be used with caution. Figure 15 shows reasonable tracking results, and the time constraint achieves correct targets disambiguation. However, this method cannot guarantee tracking stability. This technique, as shown in the results, can drift away from the tracked object, as seen for the car at time s, where the error starts to increase. Ignoring the spatial dimension shows its limits. Due to the time pattern contained in the event rates, if an event is assigned to the incorrect target, this will imply irreversible errors in the computation of interevent timings. The temporal constraint provides valuable information for segmenting the tracked targets, but it should be used in combination with the spatial constraint.
Spatiotemporal constraint. Planes are fitted on the set of events occurring inside a time window for which constant speeds of tracked cars are assumed (3 s in this case). Figure 15, frames f1 to f3, are image patches where correct trackings are achieved. The average errors shown in Figure 15 are for the van and for the car. This solution combining both space and time constraints provides the best disambiguation between both vehicles as the size of the small object is not affected by the larger one as in the previous cases. Figure 16 shows the events generated by the van and the car in the spatiotemporal space. Two planes, P1 (Van) and P2 (Car), with their respective normals and , are fitted to the events. The spatiotemporal event-to-plane distance adds more constraints, supporting the correct assignment of the event to the tracked objects, especially in case of occlusions. In cases where there are no occlusions, the use of the spatiotemporal constraint allows decreasing the tracking errors (see Figure 15 f). This is due to the fact that local spatiotemporal planes constrain strongly the neighborhood by discarding possible wrong assignments and more robustness to noise. It is interesting to note that errors are more equally shared between the van and the car as a result of the more reliable event attribution.
The exclusive use of spatial data ignoring time logically implies less available information. The dynamic content of visual information is then lost. On the other hand, time alone is not sufficient to overcome all ambiguities. The events rate provides incomplete information on the scene. The combination of both space and time then likely leads to more accurate results. A high temporal resolution combined with spatial information allows us to determine when and where things happen. More important, it allows us to define an accurate local space-time subset that necessarily contains the events generated by a target. This valuable geometric constraint is valid on a short temporal interval depending on the velocities of targets.
In principle, the determination of local spatiotemporal subsets using images is possible, but their low temporal resolution will provide less accurate results in local plane fitting, thus producing less accurate estimations. Increasing the frame rate would overcome this limitation but at the cost of losing low computation processing and subsequently at the cost of not being able to reach real-time requirement.
4.6 Pattern-Based Tracking with Camera Motion
When the camera is in motion, events are also generated on the background. One gets computationally free edge detection on both the moving object and the moving background. Figure 17 shows the typical distance distribution of object events to the plane and the distance distribution of background events to the plane. As one can see, the object events’ distances satisfy a much narrower distribution than the background events. Therefore, this shows that the mean plane is a sufficiently discriminating criterion to separate tracked object events from the nontracked ones, including those from the background. All events not consistent with the tracked object or equivalently “too far” from the plane are simply discarded.
Two sequences have been tested using the full event-based approach combining both space and time. A “star” target is moved in an indoor environment, while the vision sensor is hand-held and moved simultaneously. Figure 18, frames a1 to a5, shows several snapshots of the scene, with tracking results shown in green. To compare results, the estimated velocity of the star is calculated and shown in Figure 18a. Note that this velocity combines the star’s motion and the camera motion. The red dot line represents the ground truth. It is obtained by manually tracking specific portions of the star in each generated event accumulation frame, while the blue curve shows the velocity estimated using the event-based tracking approach. The average error between the tracked and the ground truth data is 6%. Although the scene is filled with complex background edges, the tracker remains robust. The second sequence is an outdoor traffic scene in which a car is tracked by a handheld moving asynchronous vision sensor. Figure 18b shows the velocity estimation results. Although difficult situations are present (e.g., when sidewalks or crosswalks are close to the car or even superimpose with it), the spatial temporal property of the algorithm is capable of solving such ambiguous cases. The average velocity estimation error is close to 15%. This average error is slightly higher because the algorithm is used outside its boundaries. The tracking quality starts to degrade around 3.5 s (see Figure 18, frame b5) when the car starts to have strong perspective changes. This problem will be dealt with in our future work, in which the algorithm should infer 3D information from the scene and from 3D models to compensate for perspective distortions. In summary, this additional experiment shows that the event-based camera, combined with an adequate spatiotemporal algorithm, is able to achieve robust pattern-based tracking even if the camera and the tracked objects are moving.
5 Discussion and Conclusions
This letter presents a novel methodology to perform pattern-based tracking. It shows that pattern tracking can be performed in real time at a lower cost without the need of the usual trade-offs involved in high frame-rate tracking. Sampling the x-y-t space-time volume of pixel data is shown to be an efficient input for pattern tracking. We have shown that high temporal accuracy offers a new alternative to determine objects’ velocities that are contained in a plane defined within the x-y-t volume of visual information. The sparse acquisition of visual data reduces the redundant information and retains relevant information of the scene’s dynamics. This allows the reduction of computational load, as shown by experiments. The event-based computation we have presented captures the asynchronous and time-continuous computation presented in biological nervous systems and allows shape tracking at an equivalent frame rate of . The process is data driven and iterative and it provides a natural robustness to occlusions.
The asynchronous method is generic enough to track robustly objects even if the camera itself is moving. This implies having prior knowledge of the shapes of the objects to track. The combination of time and space information provides an efficient constraint to disambiguate moving targets. This allows us to discard events that are not consistent with the targets in shape and in motion. This strategy is systematic; it can be applied to all events, from the background and from the targets. This work has produced one of the first applications showing the viability of using an event-based camera for machine vision in cluttered environments, with the camera itself moving in the scene. We believe that this work will open a new route for frame-free machine vision, allowing fast and low-cost computations that are two essential properties of several robotics tasks.
We are grateful to both the CapoCaccia Cognitive Neuromorphic Engineering Workshop and the NSF Telluride Neuromorphic Cognition workshops. This work was supported by LABEX LIFESENSES (ANR-10-LABX-65 and ANR-11-IDEX-0004-02)