基于深度强化学习的监控视频车辆型号精细识别研究  

Research on Fine Model Identification of Surveillance Video Vehicles Based on Deep Reinforcement Learning

作　　者：万淑慧 WAN Shuhui(Department of Network Security Henan Police College,Zhengzhou 450046,China)

机构地区：[1]河南警察学院网络安全系,河南郑州450046

出　　处：《传感器世界》2023年第12期29-33,共5页Sensor World

基　　金：2023年度河南警察学院院级项目(No.HNJY-2023-30)。

摘　　要：为了解决传统监控视频车辆型号精细识别存在误差较大的问题,提出了一种新的基于深度强化学习的监控视频车辆型号精细识别方法。通过聚类分析方法提取监控视频关键帧图像,并对关键帧图像进行最临近插值处理,对插值后图像做一次开运算与一次闭运算,即可获得图像中车辆的大致轮廓,得到车辆识别数据集。通过深度强化学习构建一个深度可分离卷积模型,输入待识别数据集进行模型的训练学习,完成监控视频车辆型号精细识别。实验结果表明,设计方法识别不同车辆型号的准确率高达95.16%,说明该方法具备较高的识别精度,对于交通管理和城市交通发展具有积极的推动作用。In order to solve the problem of large errors in fine recognition of vehicle models in traditional surveillance videos,a new method based on deep reinforcement learning for fine recognition of vehicle models in surveillance videos is proposed.Extracting and extracting keyframe images from surveillance videos using clustering analysis methods,and performing nearest neighbor interpolation on the keyframe images.Performing one open and one close operation on the interpolated images can obtain the approximate contours of vehicles in the images,resulting in a vehicle recognition dataset.Construct a deep separable convolutional model through deep reinforcement learning,input the dataset to be recognized for model training and learning,and complete fine recognition of vehicle models in surveillance videos.The experimental results show that the accuracy of the design method in identifying different vehicle models is as high as 95.16%,indicating that the method has high recognition accuracy and has a positive promoting effect on traffic management and urban traffic development.

关 键 词：深度学习 强化学习 监控视频 车辆型号 精细识别 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]