摘要
多域卷积神经网络(MDNet)在目标出现背景杂乱、目标遮挡、尺度变化和旋转形变时,存在跟踪精度不高、成功率下降的问题。针对此问题,提出一种融合高效通道注意力机制和可变形卷积的目标跟踪算法(CAMDNet)。通过在MDNet网络中引入高效通道注意力机制有效学习特征通道之间的相关性,进行特征筛选,增强网络特征表达能力,并引入可变形卷积以提高模型对尺度变化的建模能力,增强网络健壮性。在视频目标跟踪基准数据集OTB50和OTB100上进行评估,并与当下较为流行的跟踪算法进行对比。实验结果表明CAMDNet算法优于其他对比算法,并且比同等实验条件下的MDNet跟踪精准率提升2.25%,跟踪成功率提升2.6%,证明CAMDNet算法能有效地提高目标跟踪性能并具有较好的鲁棒性。
Multi-domain convolution neural network(MDNet)has the problems of low tracking accuracy and low success rate when the target appears background clutter,target occlusion,scale change,and rotation deformation.To solve the problems,a target tracking algorithm(CAMDNet)combining efficient channel attention mechanism and deformable ConvNets is proposed.In the MDNet network,an efficient channel attention mechanism was introduced to effectively learn the correlation between feature channels,and feature screening was carried out to enhance the network feature expression ability.Deformable ConvNets was introduced to improve the modeling ability of the model for scale change and enhance the robustness of the network.We evaluated the algorithm on the video target tracking benchmark datasets OTB50 and OTB100,and compared it with the currently popular tracking algorithm.The experimental results show that the CAMDNet algorithm is superior to other comparison algorithms,and the tracking accuracy is improved by 2.25%and the tracking success rate is improved by 2.6%compared with MDNet under the same experimental conditions.It proves that the CAMDNet algorithm can effectively improve the target tracking performance and has better robustness.
作者
贾金露
姚自强
赵玉卿
钱育蓉
Jia Jinlu;Yao Ziqiang;Zhao Yuqing;Qian Yurong(College of Software,Xinjiang University,Urumqi 830046,Xinjiang,China;Key Laboratory of Software Engineering,Xinjiang University,Urumqi 830046,Xinjiang,China;Key Laboratory of Signal Detection and Processing in Xinjiang Uygur Autonomous Region,Urumqi 830046,Xinjiang,China)
出处
《计算机应用与软件》
北大核心
2023年第9期109-116,共8页
Computer Applications and Software
基金
国家自然科学基金项目(U1803261,61966035)
自治区研究生创新项目(XJ2020G074)。
关键词
高效通道注意力机制
视频目标跟踪
可变形卷积
多域卷积神经网络
深度学习
Efficient attention mechanism
Video target tracking
Deformable convolution
Multi-domain convolutional neural network
Deep learning
