基于深度学习的实时人流统计方法研究

Research on Real-Time Statistics of People Flow Based on Deep Learnning

在深度学习出现以前,传统的人流统计方法往往难以适应复杂多变的场景,并且效率低下,难以满足实时统计要求。为使得这一工作更加趋向于智能化,本文通过将改进的轻量级目标检测网络YOLOv3-MP(YOLOv3-MobileNet-Person)与基于检测的多目标跟踪算法DeepSORT(Simple Online and Realtime Tracking)融合,设计了一种基于深度学习的YOLO-MP-DeepSORT实时人流统计方法。实验表明,本文所给的方法对行人的识别率最终高达91.67%,运行速率达到29.81 fps,功能稳定,满足实时性要求。

Before the advent of deep learning,traditional people counting methods were often difficult to adapt to complex and changeable scenarios,and were inefficient and difficult to meet real-time statistics requirements.In the work,by integrating the improved lightweight object detection network YOLOv3-MP(YOLOv3-MobileNet-Person)with the tracking by detection multi-target tracking algorithm DeepSORT(Simple Online and Realtime Tracking),A YOLO-MP-DeepSORT real-time flow statistics method based on deep learning is designed to make this work more intelligent.The experiments show that the method presented in the work can make people counting accuracy rate up to 91.67%and the Frames Per Second(FPS)up to 29.81,which has stable functions and meets real-time requirements.