Research On Road Multi-object Detection Algorithm Based On Convolutional Neural Network

With the rapid development of the Internet of Things and artificial intelligence,autonomous driving technology has attracted widespread attention from academia and industry around the world.Among them,road multi-object detection is one of important implementation technologies for autonomous driving.Due to the complex and varied traffic scenes,the detected objects often have different morphological,color,bright and occluded factors,which make the multi-object detection algorithm based on traditional statistical learning unable to meet the need of automatic driving.Researching high-performance road multi-object detection algorithms have become an urgent and challenging task.Because Convolutional Neural Network possesses the outstanding ability of self-adaptation,self-learning and fast processing operation,which provides a novel idea for road multi-object detection.Therefore,based on the theory and method of Convolutional Neural Network,the research on the road multi-object detection algorithms based on Convolutional Neural Network are carried out in this paper.The main research contents are as follows:(1)Research on vehicle object detection algorithm based on Faster Region-based Convolutional Neural Network(Faster R-CNN).Firstly,the vehicle object detection system framework is designed.The framework structure and the loss function of Faster R-CNN are deeply studied.Then the vehicle data set is created and expanded by the data enhancement technology.The model parameters of Faster R-CNN are fine-tuned and trained at the created vehicle data set,which analyzed the effect of network parameters on the performance of object detection.Finally,the experimental results show that when Faster R-CNN is used for vehicle object detection,some problems are observed during the working process,such as low detection accuracy,slow detection speed and difficulty in detecting small objects.(2)Aiming at the low detection accuracy,slow detection speed and difficulty in detecting small objects,a modified Faster R-CNN vehicle object detection algorithm is proposed.The Inception structure is introduced to extract the more detailed features of vehicles.According to the characteristics of the vehicle images,such as complicated background and object variable,the size parameters of anchor is redesigned.An Accurate Vehicle Region Network(AVRN)and a Vehicle Attribute Learning Network(VALN)are also constructed.To improve the speed of the vehicle object detection,AVRN and VALN models are alternately optimized and jointly trained.Experimental results show that the performance of the improved Faster R-CNN algorithm is significantly enhanced compared with the existing vehicle detection algorithm in terms of detection accuracy and speed.Comparing with the state of the art of Faster R-CNN detection algorithm,the modified algorithm increased the mean Accurate Precision(mAP)of detection results by 0.19 and the detection speed increased by 1/3.(3)A Fast and Accurate Road Multi-Object Detection(FAROD)algorithm based on Faster R-CNN is constructed,which aims to solve the problems of low recognition,poor detection performance and slow detection speed of small objects in the road multi-objects detection(vehicles,pedestrians and cycling)in the actual traffic scenes.FAROD includes an Accurate Object Proposal Network(AOPN)and an Object Attribute Learning Network(OALN).In order to improve the detection performance of the network,a deconvolution structure is introduced,and the loss function of the AOPN and OALN networks is designed.To speed up the algorithm,AOPN and OALN alternately optimizing and jointly training.The experimental results show that the mean Accurate Precision(mAP)of the test was 0.15 and the detection speed increased by 3 fps higher than that of the most advanced object detection algorithm Faster R-CNN,respectively.