基于跟踪—关联模块的多目标跟踪方法研究
A Tracking-Association Module-Based Study of Multi-Object Tracking Methods

多目标跟踪面临的最大挑战是身份转换问题，其由目标物体间的相互遮挡造成.针对该问题，提出一种基于跟踪模型和关联模型的多目标跟踪方法.首先在跟踪模块针对每一个跟踪个体采用粒子滤波器还原其各自轨迹片段，并计算可信因子评估遮挡程度；然后在关联模块，将人体分割为头部、躯干和腿三部分，将人的面貌分为前侧和背侧两种，利用HSV颜色直方图方法提取各部分特征描述符，利用K最近邻方法探测个体之间的匹配程度，进行再次识别以实现轨迹片段的融合.实验结果表明，同传统的方法相比，提出的算法可有效避免由于遮挡引起的身份转换问题，且目标检测准确率有较大提高，检测准确率达到90.5%.
The most important challenge faced by algorithms designed for multi-object tracking is the identity switches due to occlusions and interactions between the same tracked objects. For the problem, the paper proposes a new multi-object tracking method based on the tracking module and the association module. Through the tracking module, the trajectory segments are recovered for each tracked individual based on the use of dedicated particle filters. In the association module, the body of the individual is segmented into three parts: heads, torso and legs, and his/her appearance is classified into two poses: front and back. Then, the HSV color histogram method is utilized to extract the feature descriptors of each part, and the K nearest neighbor algorithm is used to detect the match degree between individuals, which is re-identified to achieve the fusion of each trajectory segment. The experiment results demonstrate that compared with the traditional method, the proposed algorithm can effectively avoid the problem of identity transformation caused by occlusion, and achieve a detection accuracy as high as 90.5%