Research On Pedestrian Detection In Traffic Scene Based On SSD Algorithm

Pedestrian detection technology is a key technology in the intelligent assisted driving system,and has been one of the research hotspots in the field of computer vision.It can effectively reduce the occurrence of traffic accidents,improve the safety of driving,and is of great significance to the safety of people's lives and property.However,due to the changeable posture,complex background environment and small pedestrian target in traffic scene,the current pedestrian detection algorithm has the problems of low detection accuracy and slow detection speed in practical application.Therefore,aiming at the pedestrian detection task in the road traffic scene and considering the need of accuracy and real-time detection,this paper studies the detection of pedestrian targets in the traffic scene based on SSD algorithm.The main work of this paper is as follows:Firstly introduces the working principle of mainstream YOLO target detection algorithm and the implementation process,the pedestrian traffic scene detection and analysis of experimental data,the algorithm testing speed faster,but for large scale change objects generalization ability is poorer,when the image is too dense target,prone to error detection,leak.Secondly,in order to solve the above problems,this paper mainly studies the SSD algorithm which also meets the real-time requirements with YOLO algorithm.This algorithm achieves the same accuracy as the two-stage method while maintaining a fast running speed.However,the traditional SSD algorithm is still not ideal for the detection of pedestrian targets in small and medium scale traffic scenes.Based on this,the paper makes the following improvements on the network structure of SSD algorithm:(1)the convolution layer is combined with sparse connection.By combining the convolution layer of the trunk network with the sparse connection in the Inception module,the convolution structure is optimized to increase the semantic information of the small and medium targets in the feature map,so as to enhance the feature extraction ability of the network.(2)the detection module adopts residual structure.When the multi-scale feature map is extracted from the SSD network,the detection module composed of residual structure is used to replace the traditional convolution kernel of 3×3 size to further extract the deep features of the feature map and improve the detection accuracy of the small pedestriantarget.(3)Introduce Focal Loss function.The Focal Loss function is introduced to design a new Loss function,and the weight of samples is adjusted by modulation factor to solve the problem of imbalance between positive and negative samples in the training process,so that the network pays more attention to the difficult negative samples with more useful information,thus speeding up the convergence of the network and making the training process more stable.Finally,YOLOv2,YOLOv3,SSD and improved SSD target detection algorithms were trained and tested on the public Caltech pedestrian data set to evaluate the performance of several algorithms for pedestrian detection in road traffic scenarios.Experimental results show that the improved SSD algorithm improves the detection accuracy and speed of pedestrian targets in small scale,thus improving the accuracy of pedestrian detection in traffic scenes and ensuring the real-time detection.Therefore,the research of pedestrian detection in road traffic scene based on SSD algorithm has important theoretical significance and wide application value.