Improved traffic sign detection algorithm for lightweight YOLOv8

摘要

Abstract In robotics and automated control systems, effective utilization of resources and real-time performance are critical to system performance. For traffic sign detection existing Convolutional Neural Networks (CNN), You Only Look Once (YOLO), etc, the former performs end-to-end detection, more senior feature representations can be learned, and performs better in complex scenarios, but it needs a lot of labeled data for training; the latter performs better in terms of accuracy and speed, and is able to quickly detect traffic signs and give location information, but requires higher computational resources. To overcome the problems of poor real-time transportation sign detection and mini-goals, a lightweight traffic sign proposed algorithm for the detection of devices based on YOLOv8. In this study, on the YOLOv8n model, instead of the normal convolution in BottleNeck, we propose to replace the Faster Implementation of CSP Bottleneck with 2 convolutions (C2f) structure in the backbone network with a partial convolution (PConv), resulting in a structure (CSPPC); Introduction of Cross-Scale Feature Fusion Module (CCFM) proposed by Real-Time Detection, Embedding, and Tracking (RT-DETR) modeling; Use Weighted Interpolation of Sequential Evidence for Intersection over Union (Wise-IoU) loss function. Modified method was trained on the road sign dataset and the number of parameters was reduced by 49.83% compared to the initial model, Mean Average Precision (mAP) increase of 3.92%, and the decrease in floating-point operations is 34.57%. Results show that YOLOv8n model meets the requirements of high detection accuracy and lightweight modeling with great advantages in robotics and automated control systems to accelerate the recognition and decision-making process and improve the response speed of the system.