Research On Object Detection In Traffic Scene Based On Deep Learning

As a very frontier and active research field,computer vision has developed into one of the most important branches of artificial intelligence.Vision-based object detection algorithms have been applied in various fields,such as intelligent video analysis systems,intelligent industrial manufacturing systems,intelligent urban traffic systems,automatic driving,etc.This thesis mainly studies the object detection algorithm of traffic scene.It is expected that the object detection algorithm based on deep learning can accelerate the development of autonomous perception technology.The work in this thesis is summarized as follows:Firstly,this thesis summarizes the traditional object detection algorithm and the deep learningbased object detection algorithm,and analyzes the performance differences and applicable scenarios of the two algorithms in detail.This thesis makes an objective summary of the research status of object detection algorithm based on deep learning,and gives a prospect of the future development trend.Secondly,in order to meet the demand of the accuracy and speed of object detection algorithm in the traffic scene,this thesis selects the fast one-stage object detection algorithm SSD(Single Shot MultiBox Detector,SSD)to study.To solve the problem of poor detection performance of SSD to difficult small objects,a deep feature fusion algorithm DFSSD(Deep Fusion based Single Shot Multi Box Detector,DFSSD)is proposed to improve the ability of SSD to detect difficult small objects.On the PASCAL VOC2007 test set,DFSSD increase the m Ap(mean Average Precision,m Ap)of original SSD300 by 3.7%.On the KITTI data set,DFSSD increase by 2% and 5% over the original SSD300 m Ap for moderate and difficult objects,respectively.Thirdly,in order to suppress the redundant features generated by DFSSD algorithm due to fusion and prevent the model from failing to pay attention to the key features,this thesis introduces the currently popular attention mechanism,namely CBAM(Convolutional Block Attention Module,CBAM).The CBAM module is improved and used in the DFSSD algorithm.The DFSSD-CBAM object detection algorithm is designed to improve the efficiency of deep feature fusion and further improve prediction accuracy of the algorithm.On the VOC2007 test set,DFSSD-CBAM increase by 2.1% over DFSSD’s m AP.Finally,in order to reduce the number of parameters of the model so that it can be deployed to the mobile platform with limited hardware resources in the traffic scene,this thesis studies the design principle of three lightweight models in detail,and the Mobilenet V2 is improved to serve as the backbone network of DFSSD.On this basis,a lightweight model E-Mobilessd is designed.E-Mobilessd reduces a large number of parameters without losing too much precision.E-Mobilessd has only about a quarter of the number of SSD’s,which effectively reduces the memory consumption and has the performance of real-time detection.