Research On Object Detection Algorithm Based On Autonomous Driving Scenario

Recently,as the development of the automotive field and artificial intelligence technology,autonomous driving technology has become one of the research hotspots at this stage.Computer vision plays a huge role in the field of autonomous driving.As one of the key algorithms,object detection provides great help for autonomous driving.However,the objects to be detected in the autonomous driving scene,including various vehicles,pedestrians,traffic lights and traffic signs,etc.,vary in size,have different degrees of occlusion,and are mostly moving objects.Therefore,it is a huge challenge to design an efficient object detection algorithm for such complex scenes.Based on the above content as the research background,this paper aims to design an efficient object detection algorithm in the autonomous driving scene based on the current hotspots and problems.The main contents are as follows:(1)Aiming at the problem of false detection and missed detection of small objects in autonomous driving scenarios,as well as the problems of false detection and missed detection,the AD-Faster-RCNN object detection algorithm is proposed based on Faster-RCNN.First use deformable convolution and spatial attention mechanism to improve the Res Net-50 backbone network to enhance the feature extraction of small objects and occluded objects;then introduce an improved feature pyramid structure to reduce the loss of features;introduce three levels at the same time The coupled detector solves the problem of IOU threshold mismatch,and applies the side border positioning for border regression;finally,Soft-NMS is used to remove redundant borders to obtain the best results.The experimental results show that AD-Faster-RCNN can effectively detect small objects and occluded objects in autonomous driving scenarios,and it is 7.7% higher than Faster-RCNN in the 8common categories of autonomous driving selected in the COCO dataset.(2)A real-time object detection algorithm(Mobile YOLO)is proposed for the problem that object detection is difficult to trade-off detection speed and detection accuracy in autonomous driving scenarios.The algorithm selects YOLOv4 as the base network and performs lightweight processing on it.First,Mobile Netv2 is introduced to replace the original feature extraction network,and depthwise separable convolution is used in the Neck and Head networks to replace part of the traditional convolution;then the improved lightweight channel attention module ECA is integrated to improve the representation ability of feature fusion;Finally,the SSH Context module is introduced into the 76x76 size detection branch to increase the receptive field and improve the detection ability of small objects.The experimental results show that the proposed Mobile YOLO algorithm can achieve an accuracy of90.7% on the KITTI data set,the number of model parameters is reduced by 52.11 M compared with the original YOLOv4 network,the model size is reduced to 1/5 of the original,and the detection speed is faster than the original network.increased by 70%.