边缘计算环境下大规模摄像机网络分布式实时目标跟踪方法研究

Due to the advantages of wide area monitoring, omnidirectional and multi-angle of view, and so on, the large-scale camera network can overcome the limitations in the single camera, such as complex background, illumination, occlusion, targets out of the field-of-view, etc. It has become a hot spot of present research. However, large-scale camera networks expand the scope of monitoring while produce massive video data at the same time, which brings great challenges to the real-time video data processing, especially in real-time target tracking. To address this problem, this project achieves real-time target tracking for large-scale camera networks in edge computing environment. Firstly, we establish the elastic task allocation model based on edge computing in order to improve the real-time performance and reduce the energy consumption of the system. Secondly, based on the above model, we study consensus-based distributed information fusion methods and the target state estimation method that integrates both apparent information and spatio-temporal information, and prove their validity and convergence. Finally, we will set up simulation environment and prototyped test-bed to verify the efficacy and efficiency of theoretical results and key techniques. This project has good scientific significance and applicable prospects by exploring distributed real-time target methods in large-scale camera networks, and provide thoughts and approaches for the establishment and development of other large-scale real-time video processing and analytics.

大规模摄像机网络由于具有监控范围大、全方位多视角等优点，能克服单摄像机容易受到复杂背景、光照变化、遮挡、目标脱离视野等局限，成为当前研究热点之一。然而其在扩大监控范围的同时也产生了海量视频数据，这给视频数据的实时处理，尤其对实时目标跟踪应用，带来了极大挑战。针对此问题，本项目提出在边缘计算环境下研究解决大规模摄像机网络实时目标跟踪问题。首先，以提高系统实时性和降低系统能耗为目标，建立基于边缘计算的弹性任务分配模型；然后基于此模型，研究基于一致性策略的分布式信息融合方法以及融合目标表观信息、时空信息的目标状态估计方法，并验证算法的有效性和收敛性；最后将建立仿真环境和原型系统验证理论成果和关键技术的有效性。本项目通过探索大规模摄像机网络中分布式实时目标跟踪方法，为其他实时视频分析应用技术提供理论基础和实践平台，因而具有很好的科学意义和应用前景。

本项目提出在边缘计算环境下研究解决大规模摄像机网络中实时目标跟踪问题。首先，针对边缘计算环境下硬件资源受限的边缘设备难以部署深度神经网络模型的问题，研究了深度神经网络压缩方法，并设计了边缘计算中面向数据流的实时任务调度算法和一种基于边缘计算的分割式神经网络模型；然后提出了相关噪声系统下的混合平均一致性滤波方法和基于分布式颜色的粒子滤波方法；最后搭建硬件原型平台，对所提的大规模摄像机网络中分布式目标跟踪系统的有效性进行验证。本项目通过研究大规模摄像机网络中分布式实时目标跟踪方法，为其他实时视频分析应用技术提供理论基础和实践平台，因而具有很好的科学意义和应用前景。