计算机与现代化

• 人工智能 • 上一篇    下一篇

一种智能视频监控系统中的行人检测方法

  

  1. (1.南京航空航天大学电子信息工程学院,江苏南京210016;2.南京航空航天大学无人机研究院,江苏南京210016)
  • 收稿日期:2019-03-22 出版日期:2019-11-15 发布日期:2019-11-15
  • 作者简介:杨磊(1993-),男,陕西宝鸡人,硕士研究生,研究方向:目标检测和识别,多目标跟踪,E-mail: yanghappylei@163.com; 王少云(1963-),男,江苏泰州人,研究员,硕士,研究方向:通信与信息系统,信号与信息处理,E-mail: nhshaoyun@nuaa.edu.cn。
  • 基金资助:
    国家重点研发计划资助项目(2017YFC0822404)

A Pedestrian Detection Method in Intelligent Video Monitoring System

  1. (1. College of Electronic Information Engineering, Nanjing University of Aeronautics and Astronautics, Nanjing 210016, China;
    2. UAV Research Institute, Nanjing University of Aeronautics and Astronautics, Nanjing 210016, China)
  • Received:2019-03-22 Online:2019-11-15 Published:2019-11-15

摘要: 在智能视频监控系统的行人检测中,目前使用的目标检测算法R-CNN和YOLO系列算法或速度较慢,无法满足实时性要求;或需要较大的GPU显存空间,难以部署。YOLOv3-tiny算法作为YOLO系列的精简版本,对设备要求较低、速度快,但该算法精度较低。本文通过调整YOLOv3-tiny算法的grid cell横纵方向数量、优化YOLOv3-tiny算法网络结构、聚类确定anchor的数量及尺寸,得到改进的YOLO-Y算法,并通过数据增强方法对训练数据集进行扩充。改进的YOLO-Y算法将mAP从90%提升到92%,Recall从95%提升到97%,检测速度达到26帧/s,占用约1 GB显存空间。实验结果表明改进的YOLO-Y算法显著提高了算法检测精度,具有实时性,且不需要太大的显存空间,满足大部分智能视频监控系统的要求。

关键词: 智能视频监控系统, 行人检测, YOLOv3-tiny, 聚类, 深度学习, 行人数据集

Abstract: In the pedestrian detection of intelligent video monitoring system, the current target detection algorithms R-CNN and YOLO series are slower, which cannot meet the real-time requirements, or require large GPU memory space, which is difficult to deploy. YOLOv3-tiny algorithm, as a simplified version of YOLO series, has less requirements for equipment and is faster, but its accuracy is low. In this paper, the number of horizontal and vertical directions of YOLOv3-tiny algorithm grid cell is adjusted, the network structure of YOLOv3-tiny algorithm is optimized, and the number and size of anchors are determined by clustering, so as to obtain the improved YOLO-Y algorithm, and expand the training data set by data enhancement method. The improved YOLO-Y algorithm improves the mAP from 90% to 92%, Recall from 95% to 97%, detection speed up to 26 frames/s, occupies about 1 GB of video memory space. The experimental results show that the improved YOLO-Y algorithm significantly improves the detection accuracy of the algorithm, has real-time performance, and does not need too much memory space to meet the requirements of most intelligent video monitoring systems.

Key words: intelligent video monitoring system, pedestrian detection, YOLOv3-tiny, clustering, deep learning, pedestrian dataset

中图分类号: