摘要
针对目标在经历遮挡与运动状态变化场景下的跟踪难点问题,通过在核相关滤波中引入时钟循环神经网络,提出一种新的判别式目标跟踪算法。利用双方向时钟循环神经网络构造置信图以鉴别运动目标是否被遮挡,采用置信图引导时钟循环神经网络的状态更新,从而优化调整后续核相关滤波器的学习过程。在VOT数据集上进行跟踪实验,并与19种主流跟踪算法比较,结果表明:该算法的各项跟踪指标均位于前列,且其跟踪性能显著优于同为相关滤波跟踪的其他算法。
Aiming at the difficult problems of target tracking in the scene of occlusion and changes of motion state,this paper proposes a new discriminant object tracking algorithm by integrating clockwork recurrent neural network into the kernel correlation filtering.The bi-directional clock recurrent neural network was used to construct a confidence map to identify whether a moving target was blocked.The confidence map was used to guide the state update of the clockwork recurrent neural network,so as to optimize and adjust the learning process of the subsequent kernel-related filters.The tracking experiments were carried out on the VOT dataset and compared with 19 mainstream track algorithms.The results show that the tracking indexes of this algorithm are in the forefront,and its performance is significantly better than other algorithms which are also correlation filtering tracking.
作者
吴刚
Wu Gang(School of Computer Engineering,Jinling Institute of Technology,Nanjing 211169,Jiangsu,China;Nanjing Innovation Centre of ITS,Nanjing 211169,Jiangsu,China)
出处
《计算机应用与软件》
北大核心
2020年第9期99-104,145,共7页
Computer Applications and Software
基金
国家自然科学基金青年科学基金项目(61801199)
南京市科委重大项目(201704002)。
关键词
目标跟踪
时钟循环神经网络
核相关滤波
遮挡
Object tracking
Clockwork recurrent neural network
Kernel correlation filtering
Occlusion
