Study On Air-to-ground Visual Multi-target Tracking Technology

Vision-based air-to-ground multi-target tracking technology has important application value and broad technology needs in precision strikes against ground targets, tracking and monitoring of surface ships, ground transportation monitoring and management, sea search and rescue and disaster relief, and many other military and civilian fields. Compared with the visual target tracking technology in fixed camera surveillance scene applications, the interested maneuvering targets in aerial image sequences are always small in size, and have significant changes in appearance, coupled with a lot of background noises, making the air-to-ground target detection and tracking a very difficult task. In addition, due to the complexity of ground scene, targets will be frequent occluded, seriously affecting the tracking continuousness and robustness.To solve these problems, this paper firstly studies and summarizes some of the existing feature extraction methods, target detection and tracking methods, and give a detail description of these algorithms. Based on these work, this paper puts the research focus on air-to-ground visual multi-target tracking technology, and proposed several new methods:1) Continued and robust air-to-ground visual target tracking method. The proposed method utilize particle filter as the basic framework. Firstly, the Gaussian kernel weighted mean hash features and color features of the target are extracted, and then the two features are adaptively fused based on the entropy computed by the probability distribution of the particles while constructing the observation model of the tracking system. Then the target model can be adaptively updated based on ORB feature matching level between the tracking result and the current target model. When the target is lost, a global search strategy by particle redistribution is used to search the target so that the target can be tracked when it reappears, which can improve the tracking continuousness and robustness.2) Autonomous air-to-ground multiple targets detection and tracking method. Firstly, features that can adequately describe the targets are extracted through feature selection and fusion, and the features are classified through cascade classification algorithm to get the strong classifier of the targets, after that the targets can be autonomously detected using the classifier. Then the detection results are associated and filtered through global optimal association algorithm based on Hungarian algorithm to achieve the purpose of tracking and recognition multiple maneuvering targets, the optimal association cost matrix simultaneously fused distance and direction information, which can improve the tracking robustness.3) Multi-UAV(unmanned aerial vehicles) cooperative visual target tracking method. The method fused the tracking result of each UAV using the tracking confidence as the fusion weights, which can largely reduce the tracking error and improve overall tracking accuracy, and can solve object occlusion problems. The method has overcame the limitations of single UAV tracking, and has greatly improved the tracking performance.Extensive experiments have been conducted on some air-to-ground aerial image sequences that captured under complicated conditions, and the results showed that:1) Continued and robust air-to-ground visual target tracking method can well adapted the significant and rapid changes of the target while tracking, and the target can continue to be tracked when it reappears, achieving a long-lasting robust tracking. Besides, the average frame rate of this tracking method is 15 fps, which can meet the real-time requirement in certain scenarios.2) Autonomous air-to-ground multiple targets detection and tracking method can accurately detect multiple maneuvering targets of interest, and can simultaneously track each target to be detected.3) Multi-UAV cooperative visual target tracking method can accurately judge the situation of target occlusion, and can effectively fused the tracking result of each UAV, which can continuously and robustly track the targets on the ground, and also has a high tracking accuracy.