Abstract:

Correlation filter-based tracking algorithms have shown excellent performance in the field of computer vision, but traditional correlation filters are prone to model drift in complex environments due to the use of fixed coefficient update strategies, and even fail to retrieve the tracked targets and lead to tracking failed. In order to make the tracking algorithm has better robustness when encountering background clutter and occlusion, this paper proposes an association tracking algorithm based on adaptive update strategy and redetection technology. The adaptive update strategy adaptively adjusts the template update coefficient according to the confidence of the tracking result to reduce the impact of the model drift. When it is determined that the tracked target is severely occluded or the tracking fails, the SVM classifier in the redetection strategy is used to redetect the tracked target to improve the error correction capability. The proposed algorithm is verified on the OTB2013 standard target tracking dataset and compared with the other five tracking algorithms. The target tracking accuracy and success rate are increased by 13.8% and 17.4%, respectively. When the target is occluded or the target field of view is lost, the algorithm can still retrieve the target and achieve stable tracking.

Key words: object tracking, correlation filter, confidence map, adaptive, recheck

摘要：

基于相关滤波器的跟踪算法在计算机视觉领域表现出了卓越的性能，但是传统相关滤波器由于采用固定系数更新策略，在复杂环境下很容易发生模型漂移甚至因无法重新找回所跟踪的目标导致跟踪失败。为了使跟踪算法在遇到背景杂波、遮挡等问题时能具有更好的鲁棒性，提出了一种基于自适应更新策略和再检测技术的关联跟踪算法。自适应更新策略根据跟踪结果的置信度，自适应调整模版更新系数，降低模型漂移所造成的影响。当判定所跟踪的目标遭受严重遮挡或者跟踪失败时，利用再检测策略中的SVM分类器对所跟踪的目标进行重新检测，提高纠错能力。所提算法在OTB2013标准目标跟踪数据集上进行验证并与其他5种跟踪算法进行比较，目标跟踪精度与成功率分别提升13.8%和17.4%。当出现目标被遮挡或者目标视野丢失等情况时，本算法仍然可以对目标进行重新找回，实现稳定地跟踪。

关键词: 目标跟踪, 相关滤波, 置信图, 自适应, 重新检测