基于SVM-LSTM的车辆跟驰行为识别与信息可信甄别

为利用智能车路协同系统内实时交互信息有效提升交通系统的安全性，提出了基于交通业务特征的交通信息可信甄别方法；重点构建了基于支持向量机(SVM)-长短时记忆(LSTM)神经网络的车辆跟驰行为识别与信息可信甄别模型，包括基于SVM的车辆跟驰行为识别模型和基于LSTM神经网络的车辆跟驰速度预测模型；设定了表征车辆行驶状态的特征向量，基于SVM的车辆跟驰行为识别模型将车辆行驶状态分为跟驰与非跟驰；对于跟驰车辆，基于LSTM神经网络的车辆跟驰速度预测模型根据其历史数据进行速度预测；SVM-LSTM信息可信甄别模型通过检验跟驰车辆的预测速度与其实际速度的差是否在合理范围来判断车辆数据的可信性，实现信息的可信甄别；采用公开数据集对提出的模型进行了训练与测试，并构建了不同异常类型和异常幅度的多个异常测试数据集，对基于SVM-LSTM神经网络的车辆跟驰行为识别与信息可信甄别模型进行了验证。研究结果表明:基于SVM的车辆跟驰行为识别模型对车辆行驶行为识别的准确率达到了99%，基于LSTM神经网络的车辆跟驰速度预测模型的跟驰速度预测精度达到了cm·s-1数量级；基于SVM-LSTM神经网络的车辆跟驰行为识别与信息可信甄别模型在正常数据测试集与多个异常数据测试集上的甄别正确率达到了97%。由此可见，提出的方法可用于路侧设备(RSUs)对车载单元(OBUs)实时信息和车载单元间实时信息的可信甄别。 

SVM-LSTM-based car-following behavior recognition and information credibility discirmination

To effectively improve the traffic system security by using the real-time interaction information in intelligent vehicle-infrastructure cooperative systems (i-VICS), a credibility discrimination approach for traffic information based on the traffic business features was proposed. In particular, the model for the car-following behavior recognition and the information credibility discrimination was built based on the support vector machine (SVM) and long short-term memory (LSTM) neural network. It was composed of the SVM-based car-following behavior recognition model and the LSTM neural network-based car-following speed prediction model. The feature vector representing the vehicle driving states was set, and the vehicle driving states were divided into the following and non-following by the SVM-based car-following behavior recognition model. For following vehicles, their speeds were predicted by the LSTM neural network-based car-following speed prediction model according to the history data. With the SVM-LSTM-based information credibility discrimination model, the credibility of vehicle data was judged by checking whether the difference between the predicted speed and the actual speed of the following vehicles was within the reasonable range, and in this way, the information credibility discrimination was achieved. The public dataset was employed to train and test the proposed models, and several abnormal test datasets of various abnormity types and abnormity amplitude were built to verify the SVM-LSTM neural network-based model for the car-following behavior recognition and the information credibility discrimination. Research results show that the vehicle driving behavior recognition accuracy of the SVM-based car-following behavior recognition model is up to 99%, and the predicted car-following speed precision with an order of magnitude of cm·s-1 can be achieved by the LSTM neural network-based car-following speed prediction model. The discrimination accuracy of the SVM-LSTM neural network-based model for the car-following behavior recognition and information credibility discrimination is up to 97% on the normal test datasets and multiple abnormal test datasets. Thus, the proposed approach can be applied for the real-time information credibility discriminations of road side units (RSUs) to on-board units (OBUs) and between OBUs. 8 tabs, 9 figs, 30 refs.