复杂交通场景中基于目标层次结构和语法模型的视频车辆检测方法研究

The traffic video surveillance system is an important part of intelligent transportation systems (ITS) and vehicle detection is a key technology in video-based traffic surveillance system. However, the complex traffic environments have brought about huge challenges to vehicle detection, including vehicle occlusion and stationary vehicle under congestion scenarios, lighting variation under bad weathers and at different times of day. By analyzing the problems of existing traffic surveillance system and exploring the new demands for traffic surveillance, this project studies vehicle detection technology from the point of view of object hierarchical structure and aims to realize the vehicle detection technology for complex traffic surveillance. The main research contents include: 1) The vehicle object is divided into a hierarchical structure, by analyzing the characteristics of traffic scenes and the vehicle itself; 2) According to the hierarchical structure, object detection grammars are designed for describing the composition of vehicle parts, measuring their structural relationship and finally realizing the fusion of vehicle parts. The vehicle detection method, which is realized in this project, can effectively adapt to stopping vehicle, partial occlusion, multiple vehicle views, multiple vehicle types and more practical problems. The detection model has universal applicability. The achievements of this project can be applied in the traffic video surveillance systems, by supporting and supplementing algorithms for intelligent transportation applications such as traffic information collection and traffic incident detection.

交通视频监控系统是智能交通系统（ITS）的重要组成部分，车辆检测是交通视频监控系统的一项关键技术。然而，复杂的交通环境为车辆检测带来了巨大挑战，比如，拥堵场景下的车辆遮挡、车辆静止，恶劣天气条件，一天中不同时段的光照变化等。本项目通过分析现有交通监控系统的问题和现实需求，从目标层次结构的角度研究车辆检测技术，以期解决复杂交通环境下的车辆检测技术难题。主要研究内容包括：1）分析交通场景和车辆目标的特性，将车辆目标划分层次结构；2）依据车辆目标层次结构，设计语法规则来描述车辆部件的组分方式和度量部件之间的结构关系，研究车辆层次部件的融合问题。依据本项目方案研究的车辆检测方法，能够有效处理车辆静止、部分遮挡、车辆多姿态、多车型等实际问题，具有普遍的适用性。本项目研究成果可应用于交通监控系统，为交通信息采集和交通事件检测等交通应用提供可靠的算法支持和有益补充。

交通视频监控系统是智能交通系统（ITS）的重要组成部分，车辆检测是交通视频监控系统的一项关键技术。针对复杂交通环境中的实际挑战，比如，拥堵场景下的车辆遮挡、车辆静止，恶劣天气条件，一天中不同时段的光照变化等，本课题研究了基于目标层次结构的车辆检测技术。提出了一种双层结构的部件划分方法，第一层为语义部件，根据车辆目标的特性进行选取；第二层为子部件，根据语义部件的部件模型自动获取。采用了车辆检测语法来建模车辆部件的组分方式以及它们之间的相互关系，检测语法包括形变、结构和成对支持向量机语法规则。同时研究了基于深度学习的车辆检测与识别模型，在当时达到了state-of-the-art的效果。研究成果在实际交通监控系统中进行了示范应用，为交通信息采集和交通事件检测等交通应用提供了一定的技术支持。