基于深度学习的车辆异常事件检测研究

Research on Abnormal Vehicle Event Detection Based on Deep Learning

基于全连接神经网络以及在高速公路场景下采集的真实车辆数据,设计一种车辆异常事件检测方法。所设计的方法利用激光雷达检测出目标与车道线信息,经过计算组成具备时空特征的输入数据,在将目标数据制作成数据集后,利用神经网络进行训练并提取输入数据的多维度特征,实现车辆异常事件检测。之后利用异常事件检测的时间信息与激光雷达和视频相机的目标融合数据,在视频数据中进行事件取图并加以验证。所述方法的总体检出率为97.11%,准确率为97.10%,相关值均高于传统事件检测算法,且算法耗时与传统事件检测算法相比更少。试验证明相应方法能更快速、更准确地识别车辆异常事件。

Based on the fully connected neural network and the real vehicle data collected in the highway,a vehicle abnormal event detection method is designed.The method uses laser radar to detect the information of targets and lane lines,and then forms input data with temporal and spatial characteristics through calculation.After the data set is made,the neural network is used for training and the multi-dimensional features of the input data are extracted to realize the detection of vehicle abnormal events.Then,the time information of abnormal event detection is fused with the target data of LiDAR and visual camera,and the event image is taken from the video data for verification.The overall detection rate and accuracy of the proposed meth-od are 97.11%and 97.10%,both of which are higher than traditional event detection algorithms,and the algorithm takes less time than traditional event detection algorithms.Experiments show that the method can identify vehicle abnormal events more quickly and accurately.