检测区域自适应调整的TLD多目标跟踪算法

A Multiple-Object Tracking Algorithm Using TLD-based Adaptive Adjustment of Detection Areas

传统的TLD目标跟踪算法由于检测区域过大导致检测时间过长,并对相似目标跟踪效果不理想且只能对单个目标快速跟踪.针对这些问题,利用双Kalman滤波加速预测的DKF检测区域优化算法构造了一种检测区域可自适应调整的多目标跟踪算法——TLD-DOMO算法.TLD-DOMO算法的多目标检测器可对各目标的潜在运动范围进行预测,使其检测区域的大小及位置自适应地调整至最佳状态,以此提升对多目标跟踪的精度及效率.此外,该方法可有效地降低多目标间的相互干扰,支持对多相似目标的同时跟踪.实验结果表明:TLD-DOMO算法在对各测试视频的多目标跟踪中,跟踪速度均有提升,加速比为1.55~2.94倍;在多相似目标跟踪中,对各目标的检测与识别效果优于原TLD算法.

Traditional tracking- learning-detection （TLD） object tracking algorithm takes long time for detection because the area to be detected is too large, and the algorithm is not satisfactory for tracking similar objects, and it is only suitable for single object tracking tasks. Therefore, an efficient TLD-detector optimization for multiple objects （TLD- DOMO） approach is proposed for tracking multiple objects, which is built on a novel algorithm named DKF （ double Kalman filter）. Detection areas are adjusted adaptively by the prediction method accelerated by double Kalman filtering operation. The multiple-object detectors of TLD-DOMO algorithm can predict potential motion range of each object to optimize the scale and position of detection areas adaptively. Thus, the accuracy and efficiency of multiple-object tracking will be improved. Moreover,the proposed method also reduces the interference among tracked objects effectively for supporting similar objects tracking. Experimental results show that the detection efficiency is improved in all test videos,and the speedup ratios are between 155% and 294% . The effect of detection and recognition of multiple objects surpasses the original TLD approach in tracking multiple similar objects.