基于区域预估与自适应分类的视觉跟踪算法

Visual Tracking Algorithm Based on Region Estimation and Adaptive Classification

针对视觉跟踪中的目标形变、部分遮挡和平面外旋转等问题,提出一种基于区域预估与自适应分类的视觉跟踪算法。该方法基于跟踪-修正-检测框架,利用Mean-Shift算法进行跟踪,并使跟踪器与检测器紧密相连,利用修正模块判断跟踪器和检测器是否需要在线更新;采用Kalman滤波器对目标潜在位置区域进行预估,避免全局扫描的繁琐流程;所提出的自适应方差分类器能够动态地调整分类器参数,增强分类器的灵活性,提高跟踪稳健性。采用OTB-2013评估基准中的视频序列进行测试,并将所提算法与其他4种具有代表性的视觉跟踪算法进行对比,实验结果表明,所提算法的稳健性和准确性均优于对比算法。

Aiming at the problems of target deformation, partial occlusion, and out-of-plane rotation in visual tracking, we propose a visual tracking method based on region estimation and adaptive classification. The method is based on tracking-learning-detection framework. Firstly, we use the Mean-Shift algorithm to realize the tracking, and the tracker is closely connected with the detector. The correction module determines whether the detector needs to be updated online or not. Secondly, we use the Kalman filter to estimate the potential location of the target in order to avoid cumbersome global scanning. Finally, the proposed adaptive variance classifier dynamically adjusts the classifier parameters, enhances the flexibility of the classifier, and improves robustness. Experiments perform on the OTB-2013 evaluation benchmark show that the robustness and accuracy of the proposed algorithm are better than those of contrastive algorithms.