基于改进YOLO的公路路网视频并发检测及应用  

Highway Network Video Concurrency Detection and Applications Based on Improved YOLO

作　　者：成玉荣 陈湘军 杜晨浩 胡海洋 CHENG Yurong;CHEN Xiangjun;DU Chenhao;HU Haiyang(School of Computer Engineering,Jiangsu University of Technology,Changzhou 213001,Jiangsu,China)

机构地区：[1]江苏理工学院计算机工程学院,江苏常州213001

出　　处：《实验室研究与探索》2020年第10期50-55,共6页Research and Exploration In Laboratory

基　　金：江苏省大学生创新创业训练计划项目(201911463009Z);常州市科技支撑计划(CE20185044);教育部产学研协同育人项目(201808002034)。

摘　　要：提出一种轻量化YOLO模型方法,应用于公路路网视频实时状态分析,解决了YOLO目标检测算法在实际应用中计算代价过大问题,并能满足多路监控视频实时并发检测的需求。首先,对YOLO的主干网络进行层次剪枝与通道剪枝,设计出更轻量的特征提取网络结构。其次,采用混合了多重感受野的Inception模块替换标准卷积模块,提升高层特征之间的空间信息交互。再次,修改模型数据结构,使用16位浮点数据存储参数,减少计算开销。最后,基于改进的YOLO训练车辆识别模型,实现基于车辆检测与追踪的路网断面流量统计、速度提取,以及拥堵事件检测,实时并发视频检测达到11路。实验结果表明,轻量化YOLO模型在相同检测精度下,并发检测性能提升1倍。A lightweight YOLO model method is proposed for real-time state analysis of highway network video.The method solves the problem of excessive computation cost of the YOLO object detection algorithm in practical applications,and meets the requirement of real-time concurrent detection of multi-channel surveillance video.Firstly,layer pruning and channel pruning are performed on YOLO backbone network to design a lighter feature extraction network structure.Secondly,the standard convolution module is replaced with an inception module mixed with multiple receptive fields to enhance the spatial information interaction between high-level features.Thirdly,in order to reduce computational overhead,16-bit floating-point data are used to store model parameters.Finally,the improved YOLO is used to train vehicle recognition model.The road network section traffic statistics,speed extraction,and congestion event detection based on vehicle detection and tracking are realized,and real-time concurrent video detection reaches 11 channels.Experimental results show that the video concurrent detection performance of the lightweight YOLO model is doubled without affecting accuracy.

关 键 词：深度神经网络 网络模型改进 车辆检测与追踪 车流量统计 车速检测 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]