基于卷积神经网络的列车实时客流检测算法

Real-time train passenger flow detection algorithm based on convolutional neural network

城市轨道交通作为城市公共交通系统的主体,在缓解交通出行压力、促进城市发展中具有十分重要的作用。科学精准的测算列车实时客流,对改善城市轨道交通客运服务水平,发展低碳、环保的绿色出行模式具有重大意义。为此,提出基于卷积神经网络的城市轨道交通列车实时客流检测算法,实现列车监控视频视野内乘客数量的实时检测。基于成都地铁1号线车载监控视频建立列车客流图像数据集,以剔除全连接层的VGG-16网络作为算法基础框架,提取输入图像的边缘、角点等浅层细节特征;将多尺度卷积层和膨胀卷积结构融合构建乘客多尺度特征感知模块,在保持图像分辨率的同时,通过不同的感受野增强尺度上下文信息的提取性能,提升网络对乘客尺度变化的鲁棒性;构建特征融合网络将网络浅层提取的细节特征与上采样后的深层语义特征嵌入融合,提升小尺度乘客目标的计数精度和特征图的信息丰富度。实验结果表明:所提算法在城市轨道交通列车场景下的检测精度得到显著提升,平均绝对误差、均方误差以及平均绝对百分比误差指标分别达到了1.1%,1.8%和5.4%,同时单张图像检测速率低于60 ms,能够满足列车客流检测的实时性需求;且算法在2个标准人群数据集ShanghaiTech以及UCF-CC-50上的计数性能也得到了不同程度的提升,拥有良好的泛化性能。

As the backbone of urban public transportation system,urban rail transit plays an important role in alleviating the pressure of traffic travel and promoting the urban development.Scientific and accurate calculation of real-time train passenger flow is of great significance to improve the passenger service level of urban rail transit and develop a low-carbon and environmental-friendly green travel mode.Therefore,a real-time train passenger flow detection algorithm for urban rail transit based on convolutional neural network was proposed to realize the real-time passenger number detection in the field of the train monitoring video.The train passenger flow dataset was established on the vehicle monitoring video of Chengdu Metro Line 1,with the VGG-16network excluding the fully connected layer as the basic network framework to extract the shallow detail features,such as the edge and corner point of the input image.The multi-scale convolutional layer and expansion convolutional structure were integrated to build the passenger multi-scale feature perception module.While maintaining the image resolution,the extraction performance of the scale context information was enhanced through different receptive fields,so as to improve the robustness of the network to the passenger scale change.The feature fusion network was constructed to fuse the detailed features extracted by the shallow network with deep semantic features after upsampling to improve the counting accuracy of small-scale passenger target and the information richness of the feature map.The experiments results show that the detection accuracy of the proposed algorithm in the urban rail transit train scenario is significantly improved,and the average absolute error,mean square error and average absolute percentage error indicators reach 1.1%,1.8% and 5.4% respectively.Meanwhile,the detection rate of a single image is less than 60 ms,which can meet the real-time requirements of train passenger flow detection,and its performance is also improved to varying degrees on the two standard population datasets Shanghai Tech and UCF-CC-50,with good generalization performance.