Research On Real-time Detection And Ranging System For Targets In Front Of Vehicles

In recent years,with the rapid development of intelligent vehicles,major universities and research institutions at home and abroad have invested a lot of research on autonomous driving systems.Auto equipped with an automatic driving system can effectively reduce the occurrence of traffic accidents,reduce traffic congestion,and improve road operation efficiency.The current automatic driving system is mainly researched from four aspects: path planning,environment perception,positioning and navigation,and decision-making control.Among them,the automatic driving system can obtain information about the surrounding environment through environmental perception,which has a significant impact on subsequent vehicle control and decision-making.Therefore,obtaining information about the surrounding environment of the vehicle efficiently and accurately is a particularly important part of the automatic driving system.On the basis of summarizing the current related research,this paper uses binocular vision cameras to collect image information,and detects,tracks and distances pedestrians and vehicles in the current urban road environment.It mainly includes the following aspects:(1)Analyze the development status of object detection technology,multi-targets tracking technology and stereo matching technology at home and abroad in recent years,and determine the key problems to be solved in this paper and the technical route of the paper.(2)Researched the principle of binocular camera imaging and distance measurement,compared various calibration methods,chose to use Zhang Zhengyou calibration method for calibration,analyzed the stereo matching method,and chose the stereo matching method to generate dense disparity map.Among them,SGM and AD were selected.-Census stereo matching algorithm was analyzed and their respective advantages and disadvantages were clarified.Finally,the SGM algorithm was selected as the stereo matching algorithm in this article.The SGM algorithm was accelerated through CUDA,so that the algorithm could meet the real-time requirements.The accuracy is verified by experiments.(3)The principle of the SSD detection algorithm is analyzed,and an improved SSD detection algorithm is proposed for the algorithm’s poor detection effect on small targets and the redundancy of network parameters.The Moblie Net series network is used to replace the original VGG backbone network for feature extraction,when Comparing improved the SSD algorithm with the SSD algorithm,it is found that not only the amount of network parameters is greatly reduced,but the running speed and accuracy of the detection algorithm are also improved.(4)In view of the possible misdetection and missed detection of the object detection algorithm,a multi-feature fusion algorithm that integrates the target appearance,structure,movement and size information is proposed.The appearance and the detection target are calculated based on the grid structure histogram.Structure information.Two types of motion models are designed to predict the target position of the trajectory,thereby obtaining motion and size information,effectively reducing the disappearance of the object bounding box,and providing stable object position information.The experimental verification is carried out on the open source datasets MOT15 and MOT17,respectively.The results show that the algorithm is improved compared with the SORT algorithm and the MDP algorithm.The detection algorithm is combined with the multi-objective tracking algorithm and verified on the road scene video.The results prove the effectiveness of the algorithm.(5)All the algorithms in this article are integrated into a target detection and ranging system,Among them,the detection algorithm uses Tensor RT for inference optimization,and then deployed to the Jeston Xavier NX embedded development board for experimental verification in different road scenes.The results show that the algorithm in this article is embedded in the embedded development board.The accuracy of the development board has decreased slightly,but it can still run in real time.