基于深度学习的动态目标跟踪研究  

Research on Dynamic Target Tracking Based on Deep Learning

作　　者：谌颃 钟贵[1] 戴华珍 张袖斌[1] 朱恺文[1] CHEN Hang;ZHONG Gui;DAI Hua-zhen;ZHANG Xiu-bin;ZHU Kai-wen(Guangzhou Vocational College of Technology&Business,Guangzhou 511442,Guangdong)

机构地区：[1]广州科技贸易职业学院,广东广州511442

出　　处：《电脑与电信》2023年第9期37-41,共5页Computer & Telecommunication

基　　金：2020年度广东省普通高校特色创新项目“基于深度学习的动态目标跟踪技术的应用研究”,项目编号:2020KTSCX292。

摘　　要：从深度学习卷积神经网络的角度出发,提出基于孪生(Siamese)网络框架的目标跟踪方案,利用“卡尔曼滤波+SiamRPN目标跟踪算法”,引入区域候选网络(RPN),将图像跟踪的视频帧转换为局部检测任务,运用卡尔曼预测模块排除干扰,运用SiamRPN神经网络提取检测帧、模板帧的重要特征,经过RPN候选网络的特征回归,得到动态目标对象的位置坐标、尺度,并对描述目标外观的特征信息进行加权融合,实现对全遮挡目标、噪声干扰目标的再次跟踪识别,提高动态目标对象的跟踪速度、跟踪准确率。From the perspective of deep learning convolutional neural networks,a target tracking scheme based on the twin(Siamese)network framework is proposed.The "Kalman filter+SiamRPN target tracking algorithm" is used,and a regional candidate network(RPN) is introduced to convert the video frames of image tracking into local detection tasks.The Kalman prediction module is used to eliminate interference,and the SiamRPN neural network is used to extract important features of detection frames and template frames.Through feature regression of RPN candidate networks,the position coordinates and scales of dynamic target objects are obtained,and the feature information describing the appearance of the target is weighted and fused to achieve the tracking and recognition of fully occluded targets and noise interference targets,improving the tracking speed and accuracy of dynamic target objects.

关 键 词：深度学习 动态目标跟踪 卡尔曼滤波 加权融合 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]