Research On Vchicle Target Detection System Based On Deep Learning

ADAS(Advanced driver assistance systems)can provide early warning to drivers and avoid traffic accidents,so they play an important role in ensuring driving safety.The accuracy and real-time of target recognition in the system are difficult points of current research.Among many target detection algorithms,the SSD(Single Shot Multi Box Detector)algorithm is more mature and has better target detection accuracy and speed.However,its ability to detect small targets is weak,and missing detection and inaccurate positioning when targets block each other.Therefore,this paper proposes improvement measures for the SSD algorithm to achieve fast and accurate detection of vehicles ahead.First of all,through the theoretical and practical research on target detection methods,binocular vision sensor was used to detect moving targets and process continuous frame video images,the distance of the target is estimated through the parallax of left and right images of binocular vision,and the network structure of the SSD target detection algorithm is improved,and the Conv3-3 layer is extracted as the small target prediction layer to realize the feature extraction of the small target and enhance the network’s ability to detect the small target.Secondly,a target detection data set is made,K-means clustering is performed on ground truth boxes of the targets in the data set,multiple numbers and proportions on the prior boxes are optimized,the training time of the model is changed,and the accuracy of the model detection is improved.Then,the loss function is optimized so that the prediction box is far awa y from the ground truth box of the adjacent target,avoiding the positioning distance caused by the mutual influence of the prediction boxes,and reducing the problem of missed detection due to non-maximum suppression filter out overlapping frames.Finally,the target detection results and the ranging results are associated and fused to realize the recognition and ranging of small targets and occluded ta rgets.Build an experimental environment for testing algorithms to verify the effective-ness of the algorithm.