基于场景辅助特征的T-S目标跟踪

针对遮挡同时目标附近出现相似目标干扰所导致的错跟问题，本文提出利用场景中辅助特征提升目标跟踪抗遮挡以及抗相似目标干扰性能。首先检测场景强特征及目标附近相似干扰，定义二者为场景辅助特征；其次，建立能够较好描述场景强特征及目标运动规律的动态模型以及相似干扰约束；最后，将场景辅助特征及目标的动态模型以粒子滤波的形式表达，提出T-S跟踪算法。采用SPEVI及OTB100数据库中若干典型测试视频，与近年来6种先进跟踪算法进行对比实验，并采用两种评价体系考量。实验结果表明，本文T-S算法对SPEVI多人脸、红外车辆的跟踪误差分别为24 pixel和8 pixel；对OTB100数据库中8种视频跟踪测试时，在重叠率阈值为0.5时的跟踪成功率为0.51，优于其它对比算法。本文T-S跟踪算法能够较好应对遮挡及相似目标干扰。

T-S tracking algorithm based on context auxiliary feature

Occlusion and objects with the same appearance as the target (known as distractors) are extremely challenging in the tracking domain, and distractors appearing around the target during occlusion tend to cause tracking distractors. To resolve this problem, improving the tracking performance under conditions of serious occlusions and distractor by exploiting the context auxiliary feature was proposed. First, the context strength feature and distractors around the target were detected. Second, a dynamic model that can describe the movements of context strength feature and target well and the constraint of distractor were built. Finally, the dynamic model of context strength feature and target were described in particle filter, and the T-S tracking algorithm was proposed. Using a challenging test video from the SPEVI and OTB100 datasets, the proposed algorithm was compared with other six highly ranked tracking algorithms. Two types of evaluation were adopted during testing. The experimental results demonstrate that the error pixel is 24 pixel and 8 pixel when T-S algorithm tracking multifaces and infrared car from SPEVI datasheet, when tracking eight testing videos from OTB100 dataset, the success rate of T-S algorithm is 0.51 when overlap threshold is 0.5 and surpass other compared algorithm. Our proposed T-S algorithm performs well when tracking targets under conditions of serious occlusions and distractors.