基于改进霍夫森林框架的多目标跟踪算法

针对单目视觉对多个相似的目标跟踪因遮挡等因素影响而失效的问题，提出一种基于改进霍夫森林框架的多目标跟踪算法。在将多目标跟踪问题归结为基于目标检测的轨迹关联过程基础上，通过引入在线学习霍夫森林框架将轨迹关联计算转化为最大后验概率（MAP）问题。通过在线采集多目标样本、提取目标外观和运动特征构建霍夫森林，进行森林训练得到轨迹关联概率，从而关联多目标轨迹；而引入低秩逼近Hankel矩阵进行轨迹校验，修复了误匹配的轨迹，改进了在线更新训练样本算法的效能。实验表明，轨迹误匹配率显著改善，能有效提高单目摄像机对多个相似目标有遮挡情况下跟踪的准确性和鲁棒性。

Multi-target tracking algorithm based on improved Hough forest framework

For the failure of similar multi-target tracking with monocular vision caused by influence factors such as occlusion, a multi-target tracking algorithm based on improved online Hough forest tracking framework was proposed. Based on that, the tracking problem could be formulated as a detection-based trajectories association process, and the association calculation was formulated as a Maximum A Posteriori (MAP) problem with online learning Hough forest framework. Through online multi-objective samples collection and appearance and motion information extraction, a Hough forest was constructed to associate multi-target trajectories by training for track association probability. Low-rank approximation Hankel Matrix was employed to correct the trajectories, which modified associated errors and improved the efficiency of online update of the training set. Experimental results show that the trajectory miss match ratio is significantly decreased by the proposed method, and tracking accuracy and robustness of the monocular vision are effectively improved for similar or inter-occlusion targets.