An algorithm for multiple object tracking.

摘要

In this work, an algorithm for multiple object tracking is presented and evaluated in the context of object occlusion. The shortcomings of previous methods in the case of object fragmentation and object interactions are also addressed.; The input of our tracking system is blobs in the scene. As preprocessing, our system utilized background subtraction that detects the blobs.; We proposed a graph-based tracking algorithm. In our algorithm, data association is performed in each frame by repeating the matching process and updating and processing the event graph and the hypothesis graph. We have used a merge-split approach to handle occlusion (interaction) between the objects. Dramatic changes in the appearance motivate tracking occluded objects by exploiting several frames and using global appearance information of objects before finalizing data association. So the graph representation of our tracking system enable us to exploit the information in later frames and link them to the previous frames by generating multiple hypotheses which results to a robust data association process. Our algorithm is enhanced with a fragmentation checking module which distinguishes splitting and fragmentation by monitoring the velocity of the blobs. This is based on the fact that fragmented blobs show coherent behavior and move with the same velocity. For continuous video stream, this algorithm is enhanced with an error correction module which has the advantage of discarding the wrong hypotheses and keeping only the correct ones to optimize the performance.; The output of our algorithm is object labeling and trajectories for tracked objects. To build the trajectory our tracking system collects the object's centroids during tracking and connects them when object left the scene.; Overall, our tracking system is able to track multiple objects and handle objects interaction and object fragmentation. This algorithm is designed to be more efficient by reducing the search space in hypothesis graph and more reliable in complicated multiple object tracking scenarios.