本项目面向像素级的多目标跟踪，将其看成是在主动轮廓跟踪框架下结合检测、分割和行人重识别多个任务驱动的问题。传统的基于矩形框定位目标的多目标跟踪方法由于定位不精准而依赖于后续复杂的数据关联手段。为了解决上述问题，本项目拟采用并行分支处理模式，将基于矩形框级和像素级两个分支统一在深度学习的框架中。针对跟踪弱监督情况(只有第一帧的标注信息)探索数据增强的新手段，提高前期检测跟踪的准确性而降低后续数据关联的复杂度。研究内容包括：(1)拥挤场景下的人体目标检测；(2)基于视频目标分割框架的多目标主动轮廓跟踪；(3)跟踪过程中的行人重识别及弱监督下的数据增强问题；(4)分层数据关联及推理研究。本项目在深度学习框架下，挖掘多个任务之间潜在的联系，将短期的时空一致性和长期的表观相似性有机结合，探索视觉多目标跟踪的新思路和新方法，对于复杂监控场景中的多目标行为认知计算具有重要的理论意义和应用价值。

This proposal will focus on pixel-level multi-object tracking (MOT). It is regarded as a combination of detection, segmentation and people re-identification for multi-task driven problems under the active contour-based tracking framework. Traditional box level-based MOT trackers suffer from very complicated data association mechanisms, due to inaccurate bounding box-based localization. To address this problem, this proposal is intended to adopt a box-level and pixel-level parallel branch processing mode to unify multi-tasks. With respect to weak supervision in the tracking problem (only the first frame with annotations is available), it aims to explore the new methods for data augmentation. This proposal mainly focuses on four research topics: (1) Pedestrian detection in crowded environments. (2) Multi-object active contour tracking based on video object segmentation framework. (3) People re-identification research during tracking process and data augmentation under weak supervision. (4) Hierarchical data association and inference research. Under the deep learning framework, this project plans to explore the potential connections among multi-tasks, effectively combine short-term spatio-temporal consistency and long-term appearance similarity, and explore new ideas and methods for visual multi-object tracking, which could benefit the research of multi-object behavior cognition for complex surveillance environments.