Detection And Recognition Of Traffic Signs Based On Improved YOLO Algorith

With the acceleration of urbanization and the development of transportation,traffic safety has increasingly attracted people’s attention.The traffic sign detection system in a vehicle can provide important guidance and warning to the driver,lowering the probability of traffic accidents.In recent years,deep learning technology has developed rapidly,and deep learning-based object detection algorithms have become a research hotspot.Among dobject detection algorithms,YOLO algorithm has become one of the most outstanding object detection algorithms due to its excellent performance in accuracy and speed.Therefore,this paper chooses the YOLOv5 algorithm as the main network for traffic sign detection and conducts research based on it.This paper proposed an improved YOLOv5 algorithm for detecting and recognizing Chinese traffic signs,aiming to address the multiple difficulties.The algorithm employs K-Means clustering and genetic algorithm to generate appropriate anchor boxes.The Bi-FPN is introduced to enhance the network’s representation ability.In addition,the GAM attention is introduced to enhance the network’s anti-interference ability.The SIoU loss function for bounding box regression is modified to improves the convergence speed.Finally,experiments are conducted on the Chinese traffic sign dataset.The result shows that compared with the standard YOLOv5 algorithm,the improved YOLO algorithm proposed in this paper achieves a 10.03% improvement in m AP,a 4.7% improvement in precision,a 2.6% improvement in recall,and a 3.48% improvement in F1 score,demonstrating the effectiveness of the algorithm in traffic sign detection.On the other hand,although the improved YOLOv5 algorithm can achieve high detection accuracy,it still faces difficulties when deployed in car-mounted computer systems with limited computational resource and storage space.Therefore,we embedded Ghost module to the backbone network of the YOLOv5 algorithm for less computational complexity.The experiments show that the improved Ghost-YOLOv5,while sacrificing2.8% of classification accuracy and 1.1% location accuracy,increases the inference speed by around 12%,reduces the network parameters by 16%,greatly reduces around 23%computational complexity,basically meeting the requirements of real-time detection.