摄像机监控网络跨视域目标跟踪

Cross-view objects tracking is the key problem in visual surveillance system using camera network with non-overlapping views, the goal of which is to obtain the spatio-temporal trajectories of objects for long-term, large-area tracking, by jointly analyzing the video data generated by multiple cameras. This project focuses on the fundamental problems in cross-view tracking (i.e., the lack of continuity in object motion, object attitude and observing condition), try to find new ways to enhance the performance of current cross-view tracking systems. We will ①Investigate methods for unsupervised learning of cross-view appearance mapping based on the Generative Adversarial Networks framework, design new structure for networks, generator and discriminator, define new optimization model for unsupervised learning, reducing the errors in similarity calculation caused by variances in observing conditions across views; ② Investigate methods for object parsing based on image grammar, design new grammar model for describing object attitude, study new framework for object parsing inference, investigate new similarity calculation method based on the parsing model, reducing the error caused by variances in object attitude; ③ Investigate arbitrary order data association methods based on low rank structural reconstruction of association matrix, try to compensate the lack of spatio-temporal continuity by cycle-consistency constraint, investigate the convex relaxation method for the reconstruction problem, design and analyze the algorithms for solving the covexified problem.

跨视域目标跟踪是摄像机网络视频监控系统的核心问题，通过对无重叠视域的多个摄像机视频数据的联合分析，获得目标运动时空轨迹，实现对运动目标的长时间大范围跟踪。本项目针对跨视域目标跟踪的根本问题（目标运动连续性缺失、目标位姿连续性缺失、观测条件连续性缺失），探索提升跨视域跟踪性能的新途径。①研究基于生成式对抗网络的跨视域外观映射无监督学习方法，设计新的网络结构以及生成器和判别器模型，定义无监督学习优化问题模型，通过外观映射抑制观测条件变化对相似度计算的影响；②研究基于图像语法的目标解析方法，设计适用于目标位姿描述的语法模型描述，建立新的目标解析推理框架，探索目标解析模型基础上的相似度计算方法，屏蔽目标位姿变化对相似度计算的影响；③研究基于关联矩阵结构化低秩重构的任意阶数据关联方法，利用回环一致性约束弥补跨视域跟踪时空约束的缺失，探索低秩重构问题的凸松弛方法，完成对凸松弛问题求解方法的设计与分析。

跨视域目标跟踪是摄像机网络视频监控系统的核心问题，通过对无重叠视域的多个摄像机视频数据的联合分析，获得目标运动时空轨迹，实现对运动目标的长时间大范围跟踪。针对跨视域目标跟踪的根本问题（目标运动连续性缺失、目标位姿连续性缺失、观测条件连续性缺失），本项目提出了行人目标静态和动态体征提取以及基于静动态体征的跨视域行人目标跟踪与再识别方法。由于人体体征是行人的固有生物属性，因此该方法从根本上避免了目标位姿与观测条件连续性的缺失对跨视域跟踪产生的影响。本项目将摄像机网络拓扑结构对目标运动的时空约束与目标自身生物特征进行融合，设计相应数据关联算法，进一步解决了目标运动连续性缺失的问题。对项目所提方法在公开数据集和自行采集的数据集上进行了大量的实验验证，结果表明，本项目所提方法相较于传统方法在各项性能指标上均有大幅提高，可用于极暗光照、行人换装以及长时间大范围等极端情况下的跨视域行人目标跟踪。本项目的研究成果可以为充分发挥摄像机监控网络在智慧城市、安全城市中的作用创造条件，为充分释放当前摄像机网络海量数据的潜在价值提供基础。