Research On Road Object Detection Algorithm Based On Improved YOLOv5s

With the rapid development of the economy and the rapid growth of vehicle population,the frequency of traffic accidents has gradually increased.With the rapid development of deep learning technology,the development of driving assistance systems and autonomous vehicle to reduce the frequency of traffic accidents has become a feasible solution.The key technology of driving assistance system and autonomous vehicle is road object detection,but the traditional road object detection model based on convolutional neural network is often large and difficult to meet the requirements deployed on the on-board computer.Therefore,this paper takes the road object as the research object,uses YOLOv5 algorithm to study it,and improves the network accuracy and network lightweight.Firstly,this paper introduces the research background and significance of road object detection,and introduces the research status of traditional object detection algorithm,deep learning-based object detection algorithm,and road object detection at home and abroad.Secondly,the composition and related basic theories of convolutional neural networks are introduced in detail,and the working principle and process of various classical convolutional neural networks such as Le Net-5 and Alex Net and deep learning-based convolutional neural networks such as Fast R-CNN,Faster R-CNN and SSD are briefly expounded.Then,the network structure and working principle of YOLOv5 model are introduced in detail,and then DS-YOLOv5 s,the road object detection model of this paper is proposed based on YOLOv5 model.The model improves the initial anchor box clustering algorithm,and replaces the original K-means clustering algorithm with the K-means++ algorithm,so that the anchor box clustering algorithm of the model can better adapt to the current road object detection needs.The SENet attention module is integrated into the model to improve the features according to the importance degree and suppress the unimportant features in the current road object detection task,so as to improve the detection accuracy of themodel.The NMS algorithm is improved,and the DIo U-NMS algorithm is usedto solve the problem that the accuracy of the original NMS algorithm will dec-rease in the process of model detection.Improve the spatial pyramid pooling module,improve the SPPF module to the SPPCSPC module,and further improve the accuracy of the detection model by adding a little more computation.By improving the anchor box clustering algorithm,adding attention mechanism,changing the NMS algorithm,and replacing the spatial pyramid pooling module,the accuracy of the improved DS-YOLOv5 s algorithm is improved by 3.8% compared with the original YOLOv5 s algorithm.Finally,the DS-YOLOv5 s model is lightweighted,and based on the DS-YOLOv5 s network with good road object detection performance,three lightweight convolutional neural networks such as Ghost Net,Shuffle Net V2 and Mobile Net V3 are used to lighten the DS-YOLOv5 s network to obtain GDS-YOLOv5 s,SDS-YOLOv5 s,and MDS-YOLOv5 s and other three lightweight road object detection models.Then,experiments are carried out on the VOC2012＿traffic dataset,and the results show that the size and calculation amount of the three lightweight models have been significantly reduced,and the GDS-YOLOv5 s network performs better than SDS-YOLOv5 s and MDS-YOLOv5 s,with a detection accuracy of 73.7% and a detection speed of 62.5 frames per second.The DS-YOLOv5 s road target detection model designed in this paper and the GDS-YOLOv5 s lightweight road target detection model designed based on the DS-YOLOv5 s model have good performance on VOC2012＿traffic dataset.