基于深度学习的快速长时视觉跟踪算法

A fast long-term visual tracking algorithm based on deep learning

在目标尺寸变化、遮挡和出视场等复杂长时视觉跟踪环境下,现有基于深度学习的视觉跟踪算法很难对目标进行实时准确的跟踪。针对该问题,提出一种快速的长时视觉跟踪算法,该算法由一个快速短时视觉跟踪算法和一个快速全局重检测模块组成。在基准算法SiamRPN中加入二阶通道与区域空间融合的注意力模块作为短时视觉跟踪算法,在保证快速性的同时,提高算法的短时视觉跟踪精确度和成功率;为使改进后的短时视觉跟踪算法具有快速的长时视觉跟踪能力,在算法中加入提出的基于模板匹配的全局重检测模块,该模块使用轻量级网络和快速的相似度判断方法,加快重检测速率。在OTB100、LaSOT、UAV20L、VOT2018-LT、VOT2020-LT等5个数据集上进行测试,实验结果表明,所提算法在长时视觉跟踪中具有优越的跟踪性能,平均速度达104帧/s。

Current deep learning-based visual tracking algorithms have difficulty tracking the target accurately in real-time in complex long-term monitoring environments including target size change,occlusion,and out-of-view.To solve this problem,a fast long-term visual tracking algorithm is proposed,which consists of a fast short-term tracking algorithm and a fast global re-detection module.First,as a short-term tracking algorithm,the attention module of second-order channel and region spatial fusion is added to the base algorithm SiamRPN.Then,in order to make the improved short-term tracking algorithm have a fast long-term tracking ability,the global re-detection module based on template matching proposed in this paper is added to the algorithm,which uses a lightweight network and fast similarity judgment method to speed up the re-detection rate.The proposed algorithm is tested on five datasets(OTB100,LaSOT,UAV20L,VOT2018-LT,and VOT2020-LT).With an average tracking speed of 104 frames per second,the experimental findings demonstrate the algorithm's outstanding long-term tracking performance.