基于图像处理的城市轨道交通监控系统设计

The Designof Urban Rail Transit Monitoring System Based on Image Processing

针对城市轨道交通监控系统中,由于多尺度特征提取精度低,导致乘客拥挤度检测准确率低的问题,提出将多列神经网络与空洞卷积相结合,构建一个基于人群计数的多列空洞卷积神经网络(MPCNet)。此网络采用CNN网络进行深度特征提取;然后以空洞空间金字塔池化(ASPP)为网络提供多尺度感受野,从而进行目标多尺度特征提取。实验结果表明,在Zhengzhou_MT数据集中,提出的MPCNet算法的MAE和MSE估计误差分别为1.5和2.1,均低于传统的MCNN算法和CSRNet算法。且在开发的拥挤度自动检测系统应用效果中,系统可实现轨道交通的可视化展示,可对检测结果进行批量计算,通过折线图现象得到历史数据。由此说明,本算法可对人群多尺度特征进行准确提取,可实现乘客拥挤度有效检测,具备实时性和有效性。

In view of the problem of low passenger crowding detection accuracy in urban rail transit monitoring system, it is proposed to combine multi-column neural network with void convolution to build a multi-column multi-hole convolutional neural network(MPCNet) based on population count. This network uses CNN network for deep feature extraction;then uses void space pyramid pooling(ASPP) to provide the network with multiscale feature extraction. Experimental results show that the proposed MAE and MSE estimation errors of the MPCNet algorithm are 1.5 and 2.1, respectively, in the Zhengzhou_MT dataset, which are lower than the conventional MCNN and CSRNet algorithms. In the application effect of the developed crowding degree automatic detection system, the system can realize the visual display of rail transit, batch calculate the detection results, and the historical data can be obtained through the line graph phenomenon. This shows that this algorithm can accurately extract the multi-scale features of the population, and can realize the effective detection of the passenger crowding degree, with real-time performance and effectiveness.