车联网环境下融合多源交通信息的车辆行为辨识与安全性评估方法

Vehicle behavior identification and safety assessment can effectively reduce traffic accidents. Most traditional methods only focus on the traffic incident detection by only one type of sensors, therefore, due to different external affecting factors, the accuracy of the vehicle behavior detection is not satisfactory, the result of safety accessment is not reliable and the initiative safety forewarning cannot be implemented. The emergence of Vehicular Ad-hoc Networks (VANET) provides a new approach for solving this problem. With the real-time and low redundant information provided by various sources in VANET, this project will adopt the theory of information fusion, pattern recognition and traffic engineering to study the above subjects. And the detailed plan can be mainly described as three steps. Firstly, the way, to sense and to detect the state of vehicle, road conditions, adjacent vehicles and obstacles, will be studied based on the both laboratory simulations and field testings. Secondly, the velocity behavior rules and vehicle trajectory dictionary will be studied and the vehicle behavior is to be identified by Rule Inference and Probabilistic Latent Semantic Analysis Model (PLSAM). Finally, the multi-source information from the VANET will be fused by the Improved D-S Evidence Theory (IDST), and the achievements of safety assessment and initiative forewarning will be implemented in real applications. Hence, the research results can be composed and introduced as the new idea as well as the new approach for solving the frontier problems in this research field, such as the precise positioning of vehicle, automatic vehicle behavior identification, initiative safety forewarning and so on. In addition, the research results can be used as the new theoretical guidance and the technical support for the developments and applications of relevant VANET products.

对车辆行为进行自动辨识和安全性评估可有效预防交通事故的发生。传统方法大多专注于利用单一传感器进行交通事件检测的研究，其对于车辆行为辨识的准确性和安全评估的可靠性易受外界因素影响，难以实现主动安全预警。车联网的出现为解决该问题提供了新思路。本课题将充分利用车联网中各类信息源提供的实时、低冗余度信息，综合运用信息融合、模式识别、交通工程学的理论和方法，结合多种仿真与试验手段，研究车辆自身状态、路面状况、邻车与障碍物的综合感知方法。在此基础上建立速度行为规则库和车辆轨迹字典，并利用正向规则推理和概率潜语义分析模型对车辆行为进行辨识。最后采用改进的D-S证据理论对车联网中处于不同层次的多源异构信息进行融合，实现车辆行为安全性评估和主动预警。课题研究成果将为车联网环境下的车辆精准定位、行为自动辨识、主动安全预警等前沿关键问题提供新思路和新方法，为车联网系列产品的开发和推广应用提供理论指引和技术支撑。

车辆行为辨识与安全预警是车联网领域的基础性研究问题，该问题的解决将有助于为交通管理部门和车辆用户提供基础性数据和智能化的服务，从而推动车联网从理论阶段向实用化阶段迈进。本课题的研究分为：车辆位姿精确感知、车辆行为自动识别、车辆安全预警、仿真与实验验证四个部分。1）课题提出了一种基于视觉的车辆高精度定位方法，运用车载高清相机拍摄的路面与行车环境图像，通过匹配图像中的角点特征来实现GPS盲区下的车辆高精度定位。同时，基于多传感器融合理论，对激光雷达、高清相机、超声波测距传感器的输出数据进行了多尺度融合，获得了在车辆周围短、中、远不同距离范围内的道路环境信息，为车辆的危险状态评估提供了数据基础。2）课题构建了“感知-处理-训练-识别”四层结构的车辆典型行为识别系统框架。其原理是对车载传感器采集的原始车辆状态数据进行有效的处理得出能够表征车辆行为的参数，并提取车辆行为参数的特征作为支持向量机模型的输入量，通过支持向量机训练出车辆行为的分类模型，最后对车辆的实时行为进行识别。3）课题采用改进的 D-S 证据理论对车联网中处于不同层次的多源异构信息以及车辆轨迹行为的辨识结果进行融合决策，建立了车辆行为安全性评估的融合结构与辨识框架，确定了车联网中与车辆碰撞事故密切相关的影响因素以及如何对当前车辆驾驶行为的安全性划分等级，构建了适合于车辆安全性等级评估的基本概率赋值函数，采用序贯式证据组合方法对证据冲突进行消减，最终实现对车联网环境下车辆潜在碰撞风险的判定。4）为了便于对课题提出的模型与算法进行有效测试，课题开展了多种形式的仿真研究和实车测试，采用SUMO+OMNET软件仿真平台模拟了车联网信息交互可靠性对车辆行为安全性评估及时性的影响，开发了一种车联网半实物仿真试验平台用于车辆行为识别方法的验证，并在试验场上进行了实车测试。课题可在车辆状态精确感知、车辆行为识别、风险预警等方面为车联网行业核心产品研发提供理论指导。