Lightweight Traffic Sign Detection Algorithm Based On Multi-Scale Feature Fusion

Currently,traffic accidents frequently happen in major cities and road congestion is becoming increasingly serious,and autonomous driving technology has the potential to solve these problems and has received widespread attention.As a key technology to ensure the safety of autonomous driving,traffic sign-based detection algorithms can help vehicles identify traffic sign information on the road in a timely manner,thus ensuring driving safety.However,in actual traffic scenarios,road traffic signs often suffer from problems such as being obscured by obstacles and confused with surrounding advertising logos.The main work of this paper is as follows:(1)In response to the false detection and omission phenomenon of the YOLOv7-tiny algorithm in detecting traffic signs,we improve the Efficient Layer Aggregation Networks(ELAN)module of YOLOv7-tiny and propose the Res ELAN feature extraction module.The module introduces the deep residual learning framework and uses 3×3 convolutions to replace the 1×1 convolutions of the main branch,which enable the model to avoid the vanishing gradient problem caused by the deepening of network layers and further improve the extraction ability of shallow feature information.The improved model is named YOLOv7-tiny＿Res ELAN.(2)To address the poor detection accuracy of the YOLOv7-tiny algorithm in detecting small-scale traffic sign targets,we improve the multi-scale feature fusion network.First,a160×160 detection layer is added to be responsible for small-scale target detection,and the remaining 80×80,40×40,20×20 detection layers are responsible for medium-scale and largescale target detection;second,the default parameters of the prior frame are modified by making the optimized prior frame sizes match the target classes of the dataset used in this paper further to improve the detection accuracy of the network model;finally,the Res ELAN-Tiny module is designed to replace the feature extraction module combination of Res ELAN+CBL,which can take into account the extraction of deep semantic information while reducing the computational effort.The improved model is named YOLOv7-tiny＿PAN+.(3)To solve the problem of the weak localization ability of the YOLOv7-tiny algorithm for obstructed traffic signs,we use the SIo U bounding box regression loss function to replace the CIo U loss function.This function speeds up the model’s convergence and improves the network’s localization ability for obstacle-obstructed traffic signs.The improved model is named YOLOv7-tiny＿SIo U.This paper uses the TT100 K dataset to conduct comparative experiments and visual analysis of the above three improvements respectively.The experimental results show that the proposed models above have improved m AP by 2.7%,5.1%,and 1.1% compared with the original YOLOv7-tiny algorithm.Finally,the comprehensive improvement algorithm YOLO-100 K,which incorporates the three improvements,is subjected to comparative experiments and visual analysis,and its m AP reaches 79.9%,which is a 5.7% improvement compared with the original algorithm;its FPS is 122,and the FPS of the original algorithm is 110,and there is also a significant improvement in the detection speed.In summary,the improvement measures proposed in this paper can effectively improve the detection accuracy of the algorithm and meet the real-time requirements of detection.