计算机工程与应用 ›› 2022, Vol. 58 ›› Issue (6): 192-199.DOI: 10.3778/j.issn.1002-8331.2109-0516

• 模式识别与人工智能 • 上一篇    下一篇

面向路侧交通监控场景的轻量车辆检测模型

郭宇阳,胡伟超,戴帅,陈艳艳   

  1. 1.中国人民公安大学 交通管理学院,北京 100038
    2.公安部道路交通安全研究中心 科研管理组,北京 100062
    3.公安部道路交通安全研究中心 交通政策规划研究室,北京 100062
    4.北京工业大学 城市交通学院,北京 100124
  • 出版日期:2022-03-15 发布日期:2022-03-15

Lightweight Vehicle Detection Model for Roadside Traffic Monitoring Scenarios

GUO Yuyang, HU Weichao, DAI Shuai, CHEN Yanyan   

  1. 1.School of Traffic Management, People’s Public Security University of China, Beijing 100038, China 
    2.Scientific Research Management Department, Road Traffic Safety Research Center of the Ministry of Public Security, Beijing 100062, China 
    3.Transportation Policy Planning Research Office, Road Traffic Safety Research Center of the Ministry of Public Security, Beijing 100062, China
    4.School of Metropolitan Transportation, Beijing University of Technology, Beijing 100124, China
  • Online:2022-03-15 Published:2022-03-15

摘要: 针对路侧交通监控场景和智能交通管控需要,提出轻量型的车辆检测算法GS-YOLO,解决现有模型检测速度慢、占用内存多的问题。GS-YOLO借鉴GhostNet思想将传统卷积分为两步,利用轻量操作增强特征,降低模型的计算量。在主干特征提取网络中引入注意力机制,对重要信息进行选择,提高模块的检测能力。另外参考SqueezeNet结构,使用Fire Module和深度可分离卷积减少模型参数,模型大小从244?MB降低到34?MB,内存占用降低了86%。使用Roofline模型对实验数据和模型实际性能进行分析,结果表明GS-YOLO的精确度(AP)达到85.55%,相比YOLOv4提升了约0.45%。由于受计算平台带宽影响,GS-YOLO在GPU上检测速度提升7.3%,但在CPU上检测速度提高了83%,更适用于算力资源不足的小型设备。

关键词: 图像处理, 目标检测, 轻量化, GhostNet, 深度可分离卷积

Abstract: To satisfy the needs of roadside traffic monitoring scenarios and intelligent traffic control, a lightweight vehicle detection model GS-YOLO is proposed to solve the problems of low detection efficiency and high memory consumption of existing models. Following the structure of GhostNet, GS-YOLO divides the vanilla convolution into two steps and uses cheap operations to enhance features and reduces the resource consumption of the model. An attention mechanism is introduced in the backbone feature extraction network to select important information and improve the detection capability of GS-YOLO. The model parameters are reduced by using Fire Module and depthwise separable convolution with reference to the SqueezeNet, and the model size is reduced from 244 MB to 34 MB, with an 86% reduction in memory usage. Using Roofline theory to analyze the experimental data and the actual performance of the model. The experimental results show that the accuracy(AP) of GS-YOLO reaches 85.55%, which is about 0.45% higher than that of YOLOv4. Due to the bandwidth impact of computing platform, GS-YOLO has a 7.3% improvement in image processing speed on GPU and 83% on CPU, which is more suitable for small devices with insufficient computing power resources.

Key words: image processing, object detection, lightweight, GhostNet, depthwise separable convolution