| Both the application of intelligent driving assistance systems and unmanned driving technology should ensure safe driving as a prerequisite.In the actual traffic environment,timely deceleration is undoubtedly an important part of safe driving.The role of warning obstacles and traffic control obstacles in the traffic environment is to guide vehicles to control their speed,provide safety protection,and slow down and pay attention to detours when encountering warning obstacles and traffic control obstacles.In response to the insufficient detection categories of common warning obstacles and traffic control obstacles in current related research,as well as the situation of missed detection and false detection of small targets in the current detection model,this thesis expands on existing public datasets and proposes an improved YOLOv5 s model to solve these two problems.The main work of this thesis is as follows:(1)In response to the insufficient detection categories of common warning obstacles and traffic control obstacles in current public datasets,this thesis selects the TT100 K dataset,which contains the most small-scale road signs,as the basis,removes categories with less than 100 instances in the dataset,and uses the Copy-Paste algorithm to perform data augmentation on the processed TT100 K to solve the problem of uneven sample distribution.Then,a self-built dataset is used to expand it into a dataset that conforms to the theme of this thesis.(2)In response to the missed detection of small targets by the YOLOv5 s model,this thesis proposes a method to improve the prediction accuracy of YOLOv5 s model for detecting small objects.Firstly,the K-means++ algorithm is used to re-cluster and obtain nine anchor box sizes that are more suitable for the dataset used in this study.Secondly,the Weighted Boxes Fusion(WBF)method is used instead of the original Non-Maximum Suppression(NMS)to obtain more accurate prediction box localization,thereby improving the accuracy of detecting small objects.(3)In response to the false detection of small targets by the YOLOv5 s model,this thesis proposes a feature fusion network that combines the high-level semantics in the highlevel network with the original bottom-layer features of large scale,to consider the highlevel semantic information and the high-resolution bottom-layer features that are beneficial for boundary box regression and target location,and removes single input nodes.The original feature information is directly transmitted to the semantic generation point through skip connections,which improves the detection accuracy of small targets and avoids missed and false detections.(4)In response to the problem of increased model parameters leading to decreased detection speed during the optimization process,this thesis uses Ghost convolution modules to replace ordinary convolution modules in the CBS and CSP modules in the YOLOv5 s backbone network,achieving lightweight goals.The introduction of Ghost convolution modules significantly reduces the model’s parameters and improves the detection speed.Comparative experiments were conducted on the expanded TT100 K dataset with classical algorithms including YOLOv5 s.The experimental results showed that the optimization method had a more significant improvement in the detection performance of medium and small targets,with mAPs of 88.4% and 89.5%,respectively.Compared with the original YOLOv5 s,the improved algorithm’s performance on medium and small targets was improved by 2.2% and 16.7%,respectively,and the overall mAP was improved by3.1%.In addition,the detection speed increased from 51.2 frames per second to 54.4 frames per second,and the model size was reduced by nearly 24%.Ablation experiments showed the effectiveness of multiple improvement points proposed in this thesis,and the improved algorithm had significant improvements in detection accuracy and speed compared with the original algorithm. |
