Tracking vehicles as groups in airborne videos
Neurocomputing
Hi-index | 0.00 |
Most surveillance systems adopt the paradigm that first detect and then track. However, it is hard to determine the most suitable threshold for detection since the surveillance view is unpredictable for an airborne platform. In contrast to the mainstream approach, we propose a novel approach allowing simultaneous detection and tracking. In the proposed scheme, the detection and tracking are not independent since tracking result is utilized to improve detection performance. First, frame difference is implemented to detect targets, which are called “coarse targets” here. Second, tracking is achieved by Markov Chain Mote Carlo (MCMC) particle filter. In this stage, “coarse targets” are refined. Thus, confirmed targets are categorized into four modes: enter, exit, add and delete. Here, “add” corresponds to targets which are not detected when it entered but detected in the current frame and “delete” corresponds to false alarm. Then, the number of “add” and “delete” targets is fed back to the detector to update the detection threshold by a function. Finally, we demonstrate experiments and provide evaluations of performance to show that false alarm rate and undetected rate in our approach are both 2 ~ 3 times lower than an independent detection and tracking scheme with a fixed detection threshold. Further more, we show that the optimal threshold is unrelated to the initial value. This indicates the feasibility of our method, which is essential in an airborne system.