基于深度学习的行人再识别及其在广域摄像机网络中的跟踪研究

Person re-identification has become a research hotspot in the field of computer vision and artificial intelligence in recent years. Due to the obvious changes in pedestrian posture and scale, and the application scene is vulnerable to interference from the application environment such as background, light, shadow, and camera parameters, there is no satisfactory solution for person re-identification and tracking technology in terms of robustness and speed. Aiming at the above problems, this project studies the person re-identification algorithm based on deep learning , as well as the influence of topology estimation of wide-area camera network on person re-identification and tracking. The specific innovative research content includes four aspects. (1) Combined with the deep learning algorithm, a multi-scale deep texture feature descriptor combining global and local features and combining color and texture features are proposed to eliminate the influence of light changes on color characteristics, which can improve the robustness of feature descriptors and the universality of algorithm application. (2) Using the measurement method based on deep learning, an improved person re-identification algorithm based on person re-identification mutual learning (DML) is proposed, which can effectively improve algorithm speed and accuracy. (3) A fusion target performance model is proposed to improve the topology estimation method of wide-area camera network based on cross-correlation functions, to make up for the lack of space-time relations caused by blind areas between camera fields of view, and to enhance cross-camera target transfer and continuous tracking performance. (4) A complete person re-identification system with practical application value is designed and can be used in wide-area camera networks to achieve continuous tracking of pedestrians in the entire video monitoring environment. Through the research of this project, it is expected to solve the related core technical problems on the basis of theory and algorithm, and to provide useful theoretical support for the in-depth research and application of intelligent video monitoring in wide-area camera network.

行人再识别技术近几年成为计算机视觉和人工智能领域的研究热点。由于行人姿态、尺度变化明显并且应用场景易受背景、光照、阴影、摄像头参数等环境的干扰，使得行人再识别与跟踪技术目前尚无在鲁棒性和速度方面都令人满意的解决方案。项目针对以上问题，研究基于深度学习的行人再识别算法以及广域摄像机网络的拓扑估计对行人再识别及跟踪的影响等。其创新性研究内容包括:（1）结合深度学习算法，提出采用全局特征和局部特征结合、颜色和纹理特征融合的多尺度深度纹理特征描述子，从整体上提升特征描述子的鲁棒性和算法应用的普适性。（2）采用基于深度学习的度量方法，提出改进的深度相互学习的行人再识别算法，有效提高算法速度和精度。（3）提出一种融合目标表现模型，对基于互相关函数的摄像机网络拓扑估计方法加以改进，弥补各摄像机视野间的盲区所导致的时空关系缺失，增强跨摄像机目标交接与持续跟踪的性能。（4）研究设计出一套具有实际应用价值的行人重识别系统，实现对行人在广域摄像机网络的持续跟踪。通过本项目的研究，期望在理论和算法基础上解决有关的核心技术问题，对广域摄像机网络智能视频监控的深入研究和应用提供有益的理论支撑。

在广域摄像机网络智能视频监控系统中，行人再识别和跟踪技术已经成为计算机视觉领域研究的热点。由于行人姿态、尺度变化明显并且应用场景受背景、光照、阴影、摄像头参数等应用环境的干扰，使得行人再识别与跟踪技术目前尚无在鲁棒性和速度方面都令人满意的解决方案。. 项目针对行人再识别与行人多目标跟踪问题进行深入研究，重点研究内容包括五个方面：（1）结合深度学习算法，提出采用全局特征和局部特征结合、颜色和纹理特征融合的多尺度深度纹理特征描述子，从而消除光照变化对颜色特征的影响，从整体上提升特征描述子的鲁棒性和算法应用的普适性；（2）采用基于深度学习的度量方法，结合生成对抗网络，改进深度神经网络架构进行行人再识别研究；（3）对实时定位及地图重建方法进行研究，将视频监控区域内环境的重建和对特定目标的定位与目标跟踪结合起来，弥补由于各摄像机视野之间的盲区所导致的时空关系的缺失，增强跨摄像机目标交接与持续跟踪的性能；（4）开展基于检测的行人多目标跟踪算法的研究，实现更高精度的跨摄像头行人多目标跟踪；（5）研究设计一套完善的并具有实际应用价值的基于行人再识别算法和Deepsort跟踪范式的智能寻人系统，快速找到目标行人对其进行实时定位和追踪。经过以上研究，最终形成广域摄像机网络中基于深度学习的行人再识别与跟踪的完整理论框架，设计完成一套基于摄像机网络的智能寻人系统并申请专利1项、软件著作权1项，发表12高质量学术论文，其中SCI三区1篇、EI检索2篇，培养硕士研究生7人。. 本项目对行人再识别及其在广域摄像机网络中跟踪的策略、方法、性能和应用领域作了一些有益的探索，取得了具有一定应用基础的研究成果，为行人再识别与跟踪系统的设计与开发提供了一些参考和指导意义，在理论和算法基础上解决有关的核心技术问题，对广域摄像机网络智能视频监控的深入研究和应用提供了有益的理论支撑。