自动驾驶场景下交通标志检测算法的改进

Implementation of Traffic Sign Detection Algorithm in Automatic Driving Scenarios Based on YOLOv3 Improvement

针对嵌入式平台实现交通标志检测时存在存储容量小、算力有限的问题,通过将YOLOv3算法的骨干网络Darknet53的层数缩减,使用深度可分离卷积将骨干网络中的标准卷积替换,及将算法的三个anchor-base检测头使用一个anchor-free检测头替换等改进途径,提出了DC-YOLOv3改进算法,使其在嵌入式设备中的运行速度大幅提高。相比于原始算法在嵌入式设备中的运行速度提高了接近三倍,并保持了较高的检测精度;同时对于小目标、密集目标上等各种不同环境变化具有较好的泛化能力。

When using an embedded platform to achieve traffic sign detection,there are problems of small storage capacity and limited computing power.Aiming at this problem,this paper proposes the DCYOLOv3 algorithm,which improves the running speed of the algorithm in embedded devices through the following three ways:(1)reducing the number of layers of the backbone Darknet53 of the YOLOv3 algorithm,(2)using deep separable convolution to replace the standard convolution in the backbone,(3)replacing the three anchorbase detection heads of the algorithm with one anchorfree detection head,which will be nearly three times faster than the original algorithm in embedded devices.While maintaining high accuracy,it has good generalization ability for small targets,dense targets and various environmental changes.