基于卷积神经网络的非法营运车辆识别

Research of illegal operation vehicle identification based on convolutional neural network

针对非法营运车辆在实际执法中遇到的识别和查处难,并且仅凭法律手段无法得到很好的解决的问题,提出了一种基于卷积神经网络的非法营运车辆识别方法。首先,分析非法营运车辆的特点,制定车辆轨迹生成规则,并通过仿真实验,随机生成包含6 000辆车的轨迹数据集,非法营运车辆和正常车辆各占一半;其次,通过实验验证,确定了适用于该研究的卷积神经网络结构;最后,将车辆轨迹数据处理成大小为112×112的二维轨迹特征图,作为卷积神经网络的输入,对非法营运车辆进行识别研究。仿真实验表明,重复训练12次后,该方法对非法营运车辆的识别正确率能达到90.75%,并且平均耗时的增幅较小。该方法将卷积神经网络运用到了新的领域,也为实际非法营运车辆的识别研究提供了新思路。

Since the illegal operation vehicle＇ s recognition and investigation encounters lots of difficulties during the practical enforcement, which cannot be well solved only by legal means, an identification method for the illegal operation vehicle based on convolutional neural network was proposed. Firstly, the characteristics of the illegal operation vehicle were analyzed and the rules of vehicle trajectory generation were formulated, and then the trajectory dataset was randomly generated through simulation experiments, which contained 6 000 vehicles and the illegal and normal operating vehicles are fifty-fifty. Secondly, through experimental verification, the appropriate Convolutional Neural Network structure for this research was determined. Finally, the vehicle trajectory data was processed into two-dimensional trajectory feature map, with the size of 112 x 112, as convolution neural network＇ s input, to discuss the identification of illegal operation vehicle. The simulation experiments show that this method＇ s recognition accuracy on illegal operation vehicle can reach 90.75% after 12 times repeated training, and the average consumed time has less growth.