基于TLD的舰船目标跟踪方法研究

Ship Target Tracking Based on Tracking-Learning-Detecting Tactics

复杂背景下进行舰船目标的跟踪时,在某些帧可能会有目标丢失。为了克服这个问题,采用联合检测-学习-跟踪的TLD算法。其过程是通过训练一种在线可更新的随机蕨分类器对目标跟踪结果进行检测,并使用一种基于时空约束的PN学习策略对分类器进行学习和更新,最后融合跟踪得到的结果对目标进行判别和确定。试验结果表明,该跟踪算法可适用于目标外形改变和遮挡的情况,鲁棒性强,识别率高,误检率低,同时实时性也较好,可以满足一般的在线跟踪系统的要求。

When warship targets are tracked in complex background, the targets loss may occur in some frames. In order to overcome the problem, a tracking-learning-detecting （TLD） algorithm is introduced. With the random ferns classifier which is trained online, the detection is performed based on the classification results. PN learning constrained by spatial and temporal features is used to update the classifier. The detection results and tracking results are fused to locate the target in each frame. Finally, experimental result shows that the TLD tracking algorithm has a high recognition rate and a low false detection rate. Benefitting from continuous learning with various target changes in each frame, the TLD algorithm is robust to target appearance changes and occlusion, and has a good real-time performance. The proposed algorithm can meet the requirements of general online tracking system.