基于深度学习的视频目标检测算法的实现  

Implementation of Video Target Detection Algorithm Based on Deep Learning

作　　者：孙洪迪[1] 贾民政[1] 杨民峰 SUN Hongdi;JIA Minzheng;YANG Minfeng(School of Information Engineering,Beijing Polytechnic College,Beijing 100042,China)

机构地区：[1]北京工业职业技术学院信息工程学院,北京100042

出　　处：《北京工业职业技术学院学报》2022年第1期16-21,共6页Journal of Beijing Polytechnic College

基　　金：2021年北京工业职业技术学院科研课题(BGY2021KY—17)。

摘　　要：由于光照变化、物体遮挡和复杂背景条件等众多因素的影响,目标检测一直是机器视觉领域最具有挑战性的问题。首先对视频目标检测算法中的孪生网络系列算法进行分析比较;然后将孪生网络与深度学习相结合,提出并构建全新的孪生网络跟踪器;最后将视频输入到设计好的孪生网络跟踪器中,通过网络对每一帧图像中物体的类别与位置进行准确地实时框选标注。分别将该算法和当前广泛应用的YOLOv3算法在OTB数据集上进行验证测试。测试数据表明:该算法的视频目标检测成功率和准确率均优于YOLOv3算法。Due to the influence of many factors such as illumination change,object occlusion and complex background conditions,target detection has always been the most challenging problem in the field of machine vision.Firstly,the twin network series algorithms of video target detection algorithms are analyzed and compared.Then,combining twin network with deep learning,a new twin network tracker is proposed and constructed.Finally,the video is input into the designed twin network tracker,and the category and position of objects in each frame image are accurately and real-time frame marked through the network.The algorithm and the widely used YOLOv3 algorithm are verified on OTB data sets.The test data show that the success rate and accuracy of the algorithm in video target detection are better than YOLOv3 algorithm.

关 键 词：视频目标检测 深度学习 孪生网络 YOLOv3 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]