基于跨域迁移学习的时空断续目标运动轨迹跟踪方法

With the demand of intelligent video surveillance systems , pedestrian trajectory tracking method for urban-scale video surveillance (i.e.,How to quickly push the trajectory of a particular pedestrian in the camera network and continuous tracking )has become one of key issues of video tracking.Because of the discontinuous of the object movements and the complex environment,tracking algorithm based on data-driven or model-driven association and deterministic assumptionsare often unattainable and ineffective. Sparse approximation classfication and closed-loop tracking framework based on cross-domain transfer learning are proposed in this project,which combines video mid-level feature extraction,probabilityconfidence detection,and lable roving probability hypothesis density methods. The project focuses on the deep-seated problem of pedestrian trajectory tracking in distributed multi-camera networks:probabilityconfidence-specific pedestrian detection criterion under strong approximate interference condition; specific pedestrian appearance model middle-level feature reinforcement learning update method; sparse approximation classification algorithm for transfer Learning.The project not only provides basic support for solving the trajectory relay tracking of specific pedestrians in the urban-scale surveillance scene but explores continuous estimation method of spatio-temporal discontinuous target in unstructured environment system.

随着智能监控应用的深入，研究城市级监控系统中特定行人运动轨迹跟踪方法（即：如何在摄像机网络中快速推送特定行人的运动轨迹并对其接力跟踪）成为视频跟踪的关键问题之一。由于行人运动的时空断续和复杂环境影响，单纯基于数据或模型的时空关联和确定性假设条件的跟踪算法时常效果不佳。本项目采用基于跨域迁移学习、稀疏逼近分类和闭环跟踪框架，结合视频中层特征抽取、标签漫游概率假设密度方法，重点研究多摄像机网络下行人持续跟踪方法的深层次问题：强近似干扰条件下概率置信度特定行人检测准则；特定行人表观模型中层特征抽取与增强学习更新；基于跨域迁移学习的稀疏逼近分类算法。通过标签漫游概率假设密度跟踪持续更新目标表观模型特征，结合跨域迁移学习和稀疏逼近分类算法实现特定行人的跨摄像头运动轨迹接力跟踪。为解决城市级监控下特定行人的接力跟踪提供应用基础支持，同时也对非结构化环境系统中时空断续目标的连续估计方法进行初步探索。

复杂动态场景下的视频目标持续跟踪问题受到环境复杂性、长期部分遮挡、跨摄像头运动以及目标姿势变化等多方面因素的影响。因此，传统的基于特征相似性的跟踪算法无法较好的应用于非结构化的未知环境，难以实现时空断续条件下对运动目标的持续跟踪。针对现存问题，本项目引入特征迁移机制和领域自适应的算法，结合视频中层语义特征抽取、标签漫游概率假设密度方法，重点研究了多摄像机视频监控网络下的行人持续跟踪方法的深层次问题。该项目的创新性旨在改进复杂动态场景和目标运动不确定性的目标状态建模，提升模型适应复杂环境及目标遮挡、交叉、消失后再出现等问题的能力，以及提升多行人检测和匹配准确度等方面着手，对基于迁移学习的跨摄像头目标持续跟踪算法进行深入研究。本项目提出四种高效的跨域学习算法：基于流行结构保持的判别式迁移学习算法、基于子空间对齐的判别式迁移学习算法、基于注意力机制与属性融合的跨域行人重识别算法和子任务解耦的无监督分类算法。通过在多个公开标准视频数据集上进行仿真实验，并在真实跨城市级监控视频数据上进行目标跟踪测试，其结果验证了所提算法的有效性。