Research On Intelligent Vehicle Road Multi Object Detection Algorithm Based On Improved YOLOv5 Model

Compared with traditional vehicles,intelligent vehicles can reduce traffic accidents caused by improper driving and effectively improve road safety.Intelligent vehicles can also reduce traffic jams and help improve traffic efficiency.Intelligent vehicles will greatly change the way of travel and life in the future,so intelligent driving technology is an important direction of the current automotive industry.The intelligent vehicle perception system is one of the most important and basic systems of the intelligent vehicle,and the image-based visual perception technology is the most widely used technology in the current perception system.Among them,the multi-target detection technology in the road scene is the main research content of the image-based visual perception technology.There are many difficulties in the image-based road multi-target detection technology,such as the occlusion problem in the road scene,the problem of large difference in target scale,and the problem of poor detection performance in severe weather such as fog.In the face of the above problems,the current universal multi-objective detection algorithms have not performed as expected.YOLOv5 is currently a general target detection algorithm with high detection accuracy,fast detection speed and wide application.Aiming at the difficult problems in road scenes,this paper proposes a multi-target detection algorithm in road scenes based on YOLOv5 algorithm:YOLOv5-Auto.Compared with YOLOv5,the speed of YOLOv5-Auto is increased by 8 FPS on the BDD100 K dataset,the detection accuracy and detection speed on the RTTS dataset are increased by 6.7% and 14 FPS respectively,and the detection accuracy and detection speed are increased by 9.5% and 11 FPS on the Synscapes-Foggy dataset.The main contributions of this paper are as follows:(1)In response to the problem of slow detection speed in general scenarios and poor detection effects on small targets,this paper first proposes the Rep-Res Ne Xt module based on structural re-parameterization technology.This module uses multiple branch structures during training to extract more target characteristics.During the reasoning,the single branch structure is used to speed up the reasoning.Secondly,this paper replaces the features of YOLOV5’s features fusion network to a deformed convolutional layer.Defended convolution can activate the connection between adjacent pixels,so as to better capture the spatial geometry in the image and express more complicated expression feature.(2)In order to improve the detection performance of the algorithm in the harsh scene of the fog,this paper generates a foggy dataset Synscapes-Foggy based on the Synscapes dataset.Pedestrians,cars,etc.Synscapes-Foggy not only has images of different time periods,but also has a corresponding clear image each fog and sky image.In addition,this article proposes a fog and sky feature enhancement module based on the spatial attention mechanism.This module gives a higher weight to the important features of the objects in the fog scene during training,and the unimportant features are ignored in order to reduce the fog.Sky’s interference during algorithm detection.(3)Experiment to the algorithm Yolov5-Auto proposed in this paper.The experimental results show that YOLOV5-Auto can detect the objects on the road more quickly and more accurately than YOLOV5.It has improved in terms of implementation,accuracy and robustness.