Research On Subway Pedestrian Detection And Recognition Algorithm Based On Multi-layer Feature Fusion

With the continuous construction and development of the city,many metro electric buses with better performance and better service functions have been put into operation.This not only meets the needs of people to travel,but also responds to the national call for green travel.In order to alleviate the pressure of ground traffic,subway lines are constantly expanding.The detection and identification of pedestrians is conducive to reducing the pressure of subway staff and providing technical support for intelligent systems such as real-time monitoring and early warning of passenger flow,which has practical research significance.The video monitoring system in the subway can not only ensure the real-time centralized monitoring function of mechanical and electrical equipment and the coordination and linkage between various systems,among which the big data can also be the data basis of subway pedestrian detection technology.However,in the actual situation,pedestrians are mostly in parallel,but the subway escalators and other places are also the focus of attention.More importantly,because pedestrians have the characteristics of both rigid and flexible,their appearance is susceptible to wear,scale,occlusion,attitude,and viewing angle.The dim light in the subway and the dense crowds make it more difficult to detect and identify tasks.In the current pedestrian detection technology,traditional detection methods will cause a large number of false detections and missed detections in subway pedestrian detection due to feature extraction methods,and deep learning theory can improve the pedestrian detection rate by digging deep features of the image.Therefore,this thesis chooses the pedestrian detection technology based on deep learning,researches on convolutional neural networks,data augmentation,feature fusion,and training data imbalance,etc.,and designs an actual subway pedestrian detection framework based on the idea of multi-layer feature fusion.Specifically,The work is as follows:(1)Investigate the target detection framework,study the deep learning framework based on the multi-scale prediction framework and make improvements.On the basis of matching multiple scales of the target with multiple sensory field regions of different sizes,integrate adjacent feature maps to achieve richer image feature information.In addition,according to the feature extraction network based on the design framework of deep convolutional network with dense connection structure,the detection accuracy of small and medium-sized targets of subway pedestrians can be further improved.(2)Optimization of the overall network framework.According to the pedestrian target size of the new subway data set,the feature source layer and the default box adjustment rate in the network are adjusted.Secondly,all training samples are used to participate in the calculation of loss instead of the original manual division of positive and negative samples for training,and the model is more focused on the samples that are difficult to classify during training,so as to improve the detection rate of dense targets in the data set.(3)Expand the subway pedestrian data set.The original dataset was collected from a Beijing subway station which contains two scenes.In order to diversify the scene,this thesis combined with specific projects,collected different scenes of Nanjing Metro Station(station hall security check point,escalators,etc.)to expand the original data set,and then experimentally explore the method of dividing the data set.The experimental results show that compared with the reference algorithm,the detection accuracy of the subway pedestrian detection algorithm proposed in the thesis is improved by 5.1% to 91.5%,and the detection speed reaches 33 FPS,which can meet the real-time detection requirements.