为提高复杂情况下目标跟踪的稳健性,提出一种自适应特征更新的目标跟踪算法。对目标提取分级深度特征和手工设计特征,通过不同线性组合方式进行多特征融合,构建多个融合特征器;对不同融合特征器进行可信度判定,选择可信度最高的融合特征作为当前帧的跟踪特征,构建位置相关滤波器,预测出当前帧的目标位置;对跟踪结果进行可靠性检测,可靠性低于阈值则启动融合特征器更新机制,加入时序信息和语义信息进行重跟踪,降低了模型的误差累积。在OTB-2013和OTB-2015数据库上进行测试,结果表明,与近年来比较流行的9种算法相比,提出的算法在快速运动、背景杂波、运动模糊、形变等复杂情况下具有较高的成功率和较好的稳健性。

To improve the robustness of object tracking in complex situations, a new algorithm based on adaptive feature updating is proposed. First, hierarchical deep and hand-crafted features are simultaneously extracted from the object, and multiple fusion feature experts are constructed through multi-feature fusion by using different linear combination methods. Second, the credibility score of each expert is computed and the highest score is selected as the tracking feature of the current frame. A position correlation filter is then constructed to predict the frame's target position. Finally, the reliability of the tracking result is detected. When this reliability is found to be lower than a certain threshold, the fusion feature updating mechanism is initiated, and the temporal and semantic informations are added to the re-track, which reduces the error accumulation of the model. The proposed algorithm is tested on OTB-2013 and OTB-2015 datasets, and the obtained results are compared with those of 9 recently developed popular algorithms. Our proposed algorithm demonstrates a higher success rate and better robustness in complex situations, such as fast motion, background clutter, motion blur, and deformation, than existing algorithms.