A Study On Fast Detection Algorithm Of Road Target Based On YOLO

The rapid growth of car ownership in our country brings great convenience to us,but also inevitably brings serious traffic problems.With the promotion of AI in various industries,new solutions have been proposed for solving traffic problems.Road target detection with vehicles,pedestrians,and traffic signs as detection objects is the basis for realizing the construction of intelligent transportation systems such as vehicle intelligent assisted driving and automatic driving.In recent years,target detection algorithms based on computer vision have developed rapidly.Many two-stage target detection algorithms have achieved high detection accuracy.However,due to occlusion,weather,lack of light,and detection of many small targets,there are still many serious problems such as missed detection and false detection.In addition,because the two-stage algorithm takes a long time to detect,it cannot meet the high requirements of real-time detection of road objects.Therefore,based on the YOLO one-stage target detection algorithm,this paper conducts research on the fast detection of road objects.The works of the paper mainly include:.(1)In view of the characteristics of the variety of traffic road targets and the variability of detection scenes,it is difficult for conventional target detection algorithms to find an ideal balance between accuracy and speed,resulting in slow detection speed and high false detection rate.We propose an improved YOLOv4 fast detection algorithm for traffic road target.First,a context exploitation method is introduced to reduce the information loss of feature map at the highest level in FPN;Secondly,the residual feature enhancement method is adopted to enhance the extraction of YOLOv4 neck convolution layer features;Finally,Aug PAN is used to improve the feature fusion method to enhance the representation ability of the feature map.Compared with other classic methods for road target detection on VOC and TT100 K datasets,it is found that the improved YOLOv4 algorithm can effectively reduce the missed detection and false detection of road targets,and significantly improve the detection accuracy and speed.(2)Aiming at the phenomenon that the detection of small objects in traffic signs is difficult and prone to false detection in scenarios such as autonomous driving,the YOLOv4-S algorithm is proposed.First,an improved OSA module is introduced to improve the backbone to improves the efficiency of the model,which utilizes the fusion of feature maps with different receptive fields to get diverse feature representation.At the same time,identity map is used to solve the problem of saturation and degradation that is prone to occur in deep models to improve model accuracy.Second,the E-PAN architecture is introduced at the neck of the model,which not only strengthens the fusion of feature maps with cross-scale connections,but also captures appropriate receptive fields for each anchor and better aggregates the initial feature pyramid.The research shows that the proposed YOLO-based road target fast detection algorithm can effectively reduce the model’s complexity and reduce the phenomenon of missed detection and false detection of road objects after the use of feature enhancement methods such as context information and cross-scale connection.In addition,in order to better adapt to small target detection scenarios,the fusion of feature maps is strengthened through the improved OSA module and the E-PAN architecture,which greatly improves the efficiency of the model and improves the problem of detection difficultly of small traffic signs.