基于部件学习和图匹配的目标跟踪

Object tracking is one of the foundation problems in computer vision. It is an interesting but challenging problem due to many annoying factors, e.g., pose changes, occlusion, illumination variation, complex backgrounds, large non-rigid motion and so on. In this project, we propose to utilize the adaptive structural parts learning method, to mitigate the influence of the pose changes and large non-rigid motion problem. Furthermore, we propose to utilize the graph model to represent the interactions between different parts to reflect the structure information of the target. Combining the target appearance and structure information, the graph matching method is adopted to infer the size and position variation of the target. Based on the single target appearance modeling and tracking method, we absorb the advantages of both detection and tracking module, i.e., tracking by detection to handle the drift problem and detection by tracking to address the miss detection problem. In this way, we construct the spatio-temporal structure graph to complete the multiple targets tracking problem. The approximate optimal solution to the global energy function with the global structure graph is achieved using the graph based optimization methods.

目标跟踪是计算机视觉中的一个基础问题，在复杂场景下，其难点问题包括姿态变化、遮挡、光照、复杂背景以及大尺度的非刚性运动等因素的影响。本课题提出一种基于自适应学习、目标部件、部件结构图模型的方法，解决复杂环境下的多目标检测与跟踪。首先，采用自适应学习方法学习部件检测，使得目标表象模型能够更好的适应姿态和大尺度非刚性运动带来的影响；其次，采用图模型描述目标表象信息和各部件之间的结构关系，并采用图匹配的方法结合，推测目标尺度和位置的变化，从而对单目标进行稳定的跟踪。为解决多目标的稳定跟踪，进一步结合检测和跟踪有用信息，通过检测来指导跟踪，通过跟踪来解决漏检，并构建时空结构图，利用基于图优化的方法优化结构图的全局能量函数，得到全局近似最优的目标关联结果，从而达到多目标稳定跟踪。

本课题针对目标跟踪问题展开研究。针对单目标跟踪问题，提出基于时空上下文信息的目标跟踪方法，提高长时间目标跟踪的稳定性；提出动态图匹配的目标表象建模和跟踪方法，提高目标跟踪的鲁棒性；提出联合目标分割和跟踪的方法，提高目标跟踪的精确性。针对多目标跟踪，提出基于超图稠密结构分析的多目标跟踪框架，提高多目标跟踪的准确性和鲁棒性；提出基于时空超图的多摄像机多目标跟踪方法，提高目标跟踪的准确度。所提出的方法均取得了同时期先进的目标跟踪性能。本课题建立和公布了一个多目标检测和跟踪数据库，推动目标跟踪领域的发展。