Research On Traffic Object Detection Based On Improved YOLOv5

With the development of artificial intelligence,intelligent cars have become a reality,and assisted driving technology is becoming an indispensable part of intelligent cars.The development of AI technology has enabled autonomous vehicle to land,which has not only brought convenience to people’s lives,but also brought potential safety hazards that deserve our attention and reflection.This article analyzes various technologies and current status related to autonomous driving through reading and researching a large number of literature both domestically and internationally.It focuses on traffic object detection technology with important research significance.In response to the urgent need for a real-time object detection algorithm with high detection accuracy and low false detection rate in traffic object detection,a traffic object detection algorithm with low miss detection rate,high detection accuracy,and real-time detection performance is proposed.The main research content is as follows.Firstly,the task of traffic object detection has high requirements for detection accuracy and recall rate.This article analyzes the performance shortcomings of YOLOv5 network on the traffic object detection dataset BDD100 K,and finds that the problem of weak small object detection ability still exists in the field of traffic object detection,and there is a phenomenon of excessive target overlap leading to missed detection and greatly reduced classification confidence.In response,FCIo U is proposed to be used＿Loss replaces CIo U＿Loss makes loss calculation more inclined to fit small target detection boxes,thereby improving detection recall.At the same time,coordinate attention mechanism is integrated into the neck structure of YOLO network to improve overall target detection accuracy.By incorporating candidate boxes to remove redundancy algorithm Matrix NMS,the overall performance of the model is improved.The experiment shows that the YOLOv5 s network proposed in this article,which combines coordinate attention mechanism and multi-dimensional optimization,detects traffic targets on BDD100 K,achieving a m AP50 of 78.3%,a recall rate of 89.1%,and a detection speed of 78 fps.Compared with the YOLOv5 original model,it improves the recall rate by 2.3%,effectively solving the problem of missed detection of small and overlapping targets in autonomous driving perception tasks.Secondly,in response to the urgent need for a lightweight object detection model in the field of traffic object detection,this article has made algorithm improvements on the baseline model YOLOv5 s.Firstly,the Mobile Net V3 network with a combination of lightweight dimensions is integrated into the Backbone structure of YOLOv5 s,and then the Bi-FPN network with bidirectional cross scale weighted features is fused to optimize the feature pyramid structure in the Neck structure.Train and analyze on the BDD100 K traffic object detection dataset,and validate the effectiveness of the lightweight traffic object detection model proposed in this paper through ablation experiments and relevant evaluation indicators.The experiment shows that the lightweight YOLOv5 s model improved in this article achieves a detection speed of102 FPS and high detection accuracy on BDD100 K,achieving excellent levels in speed,memory consumption,and floating-point operations per second.Compared to the original model,the detection speed is improved by 20 fps.