Integrated object detection and tracking by multiple hypothesis analysis

In this paper, we describe a novel multi-object tracking technique that integrates object detection into the object tracking process and solves the tracking problem by finding the globally optimized object trajectories through the multiple hypothesis analysis. The detection module recognizes the target objects in each frame of the video stream. The tracking module accumulates the detection results in a graph-like structure and maintains multiple hypotheses of objects trajectories. The hypotheses are ranked by their likelihoods which are computed over a sufficient number of frames, and the most likely hypothesis is used to generate the object tracking result. At the same time, the tracking module gives feedbacks to the object detection module, which are predictions of object locations in subsequent frames. Through such tight integration of the object detection and tracking, as well as the global optimization of object trajectories, we have accomplished not only robust and efficient object tracking, but also the ability to deal with occlusions, irregular object motions, changing appearances, etc. which are the challenging problems for most traditional tracking methods.