Research On Multi-target Detection Algorithm In Traffic Scene Based On Deep Learning

The research and development of the assisted driving system and autonomous driving system make the object detection task in traffic scene become an important branch in the field of current object detection.Its research content includes obstacle detection in the road,pedestrian detection,vehicle recognition and lane detection.The disturbing factors in the traffic scene include mutual occlusion between targets,uneven light and darkness,and numerous small targets.The complex traffic environment makes most object detection algorithms have poor detection performance for small targets and occlusion.At present,compared with the traditional object detection methods based on image processing,most of the object detection methods based on deep learning have made great breakthroughs in detection accuracy and real-time performance.And this kind of method is increasingly being applied to intelligent transportation systems.This paper studies the basic theory and implementation method of the traffic scene multi-objective detection system based on deep learning.Focusing on SSD,the key technologies of data set clustering analysis,basic network building and multi-scale feature fusion are studied.In response to the problem that the VGG-16,the basic network of the SSD,has weak feature extraction capabilities and large parameters,an object detection algorithm D-SSD based on Dense Net is proposed.Firstly,the feature extraction network Dense Net-S-32-1 is redesigned to replace the SSD backbone network VGG-16,which reduces the amount of parameters and calculations of the model on the premise of ensuring the detection accuracy.Secondly,K-means clustering analysis is performed on the ground truth box in KITTI traffic scene data set,and the size and aspect ratio of the default box in D-SSD model are redivided to further improve algorithm performance.Finally,the usefulness of the design components of D-SSD is analyzed through comparative experiments.The experimental results demonstrate that the detection performance of the D-SSD is improved,and the parameters of the D-SSD model are only 1/8 of that of SSD,and the detection speed can also be real-time.In view of the difficulty of small target detection in traffic scene and the mutual occlusion of the detected targets,an object detection model DF-SSD based on Dense Net and feature fusion is proposed.Firstly,Dense Net-S-32-1 is used as the backbone network of DF-SSD.Secondly,multi-scale feature fusion module is designed by using atrous convolution and deconvolution to enhance the feature complementarity between feature layers of DF-SSD.Finally,residual prediction module is constructed for each predicted layer.Compared with the SSD,DF-SSD can improve the detection accuracy to 87.3%m AP on KITTI dataset.Moreover,the parameter quantity of DF-SSD is only 1/6 of that of SSD,and the detection speed can reach real-time,while improving the SSD’s detection performance for small targets.Based on the research of the above object detection algorithms,a prototype system of traffic scene multi-target detection based on deep learning is completed.The system adopts B/S architecture,takes DF-SSD as the object detection model,and deploys the trained DF-SSD model on Windows by using the web application framework Flask.The image to be detected can be uploaded through the browser,and the result of the image detection can be returned after evaluation by the server and showed in the browser.As the DF-SSD object detection model has smaller parameters,it can also be deployed in mobile terminals,auxiliary driving systems,and automatic driving systems with limited computing resources.