基于多模式人群运动模型的在线多目标跟踪

In recent years, due to the public security issues, the visual tracking and analysis of the crowd motions in public arouses the interest of researchers. To capture the crowd motion, we aim at investigating the multi-target tracking in the crowded surveillance video. As one of the classic research problems in computer vision, the goal of multi-target tracking is to capture the trajectories of multiple objects simultaneously from a video sequence. However, it is more difficult to track crowd in surveillance videos, due to the high crowd density. Besides, for the public security surveillance, it is more informative to perform online tracking for the large scale crowds, but it makes the problem more difficult as well. In this project, we are going to leverage multi-modal crowd motion models to boost crowd tracking, which includes extracting global motion patterns from the captured data in crowd scenes, modeling the local interactions among multiple persons (i.e., collision avoidance and grouping models), and predicting the pedestrian orientations based on its appearance and pose. The multi-modal crowd motion models will improve the issues in crowd tracking such as the occlusion among persons and the low-resolution individual observations.

近几年由于社会上频发的公共安全问题，通过监控视频跟踪分析公共场所中的群体活动引起人们的极大关注。为了获取群体运动的数据，本项目将研究针对监控视频中人群的多目标跟踪。多目标跟踪作为计算机视觉中的传统问题之一，旨在从视频中同时捕获多个目标的运动轨迹。然而，由于监控视频中的人群较为密集，因此人群跟踪的难度更大，而且监控公共安全往往需要对大规模人群进行在线跟踪，这也增加额外的难度。本项目将利用多模式人群运动模型对在线多目标跟踪提供辅助。其中，多模式人群运动模型包括：从人群运动场景的运动数据中识别全局人群运动模式，多行人之间的局部交互模型（避障和群组模型）以及基于行人外观姿态的行人朝向模型。这些多模式人群运动模型将针对在线人群跟踪中常见的遮挡、目标行人分辨率较低等问题进行改善。

近几年由于社会上频发的公共安全问题，通过监控视频跟踪分析公共场所中的群体活动引起人们的极大关注。为了获取群体运动的数据，本项目研究了针对监控视频中人群的多目标跟踪。多目标跟踪作为计算机视觉中的传统问题之一，旨在从视频中同时捕获多个目标的运动轨迹。然而，由于监控视频中的人群较为密集，因此人群跟踪的难度更大，而且监控公共安全往往需要对大规模人群进行在线跟踪，这也增加额外的难度。本项目利用了多模式人群运动模型对人群行为进行分析，并针对人群跟踪问题中常见的遮挡、行人个体姿态变化等问题进行改善。项目执行过程中，完成学术论文15篇，发表期刊论文12篇，其中SCI收录9篇（含6篇中科院分区一区论文），发表国际会议论文3篇，均为CCF-A类会议论文，项目执行期间培养硕士研究生8名。