Computer Engineering and Applications ›› 2024, Vol. 60 ›› Issue (13): 361-368.DOI: 10.3778/j.issn.1002-8331.2305-0091

• Engineering and Applications • Previous Articles    

Lightweight Traffic Signal Light Detection Algorithm for Embedded Terminal

YANG Yongbo, LI Dong, FANG Jiandong, DONG Xiang, LI Yiwei   

  1. Key Laboratory of Perception Technology and Intelligent System of Inner Mongolia, College of Information Engineering, Inner Mongolia University of Technology, Hohhot 010051, China
  • Online:2024-07-01 Published:2024-07-01

面向嵌入式端的轻量级交通信号灯检测算法

杨永波,李栋,房建东,董祥,李毅伟   

  1. 内蒙古工业大学 信息工程学院 内蒙古自治区感知技术与智能系统重点实验室,呼和浩特 010051

Abstract: In view of the existing traffic signal detection algorithms, which require a large amount of computation and model, are difficult to deploy at the embedded end, and are difficult to detect long-distance traffic signals with a high miss rate, this paper designs a lightweight traffic signal detection algorithm for the embedded end, which aims at lightweight and real-time requirements. GhostNet network Ghost module and Ghost bottleneck layer structure are adopted to reduce the number of model parameters and improve the detection speed. To solve the feature similarity problem, the weighted bidirectional feature pyramid network is used to make the algorithm more sensitive to the target. The dense void space pyramid pool is used to optimize the extraction of global context information. Aiming at the problem of small target recognition, the information extraction of small target is enhanced through the improvement of multi-scale detection. Finally, the learning ability of the model is improved through knowledge distillation, and the detection performance is improved. Experimental results show that the recognition accuracy and recall rate of traffic lights by this detection algorithm reach 97.0% and 99% respectively, 2.7 and 3 percentage points higher than that of YOLOv5s. The model size is reduced to 8.06 MB, 58% of that of YOLOv5s, and the recognition rate is increased from 51 frames per second to 56 frames per second. Through testing in the embedded terminal, the improved algorithm can recognize the traffic signal in real time and accurately.

Key words: object detection, ?lightweight, ?GhostNet, ?knowledge of distillation, dense atrous spatial pyramid pooling

摘要: 针对现有交通信号灯检测算法计算量和模型大,嵌入式端部署难,且对远距离交通信号灯的检测难度大,漏检率高等问题,设计了一种面向嵌入式端的轻量级交通信号灯检测算法,针对轻量化和实时性要求,采用GhostNet网络Ghost模块和Ghost瓶颈层结构,减少了模型参数量,提升了检测速度;针对特征相似问题,采用加权双向特征金字塔网络结构,使得算法对目标更敏感;使用密集空洞空间金字塔池化,优化全局上下文信息的提取;针对小目标识别问题,通过多尺度检测的改进,增强对小目标的信息提取;通过知识蒸馏,提升模型学习能力,进而提高检测性能。实验结果表明,该检测算法对交通信号灯的识别精度达到了97.0%,召回率达到了99%,较YOLOv5s算法分别提高了2.7和3个百分点,模型大小减小到8.06?MB,是YOLOv5s的58%,识别速率从51帧每秒提升到56帧每秒,通过在嵌入式端的测试,改进后算法对远距离下的交通信号灯能够实时准确地识别。

关键词: 目标检测, 轻量级, GhostNet, 知识蒸馏, 密集空洞空间金字塔池化