用于通用目标跟踪的图记忆跟踪器

Graph Memory Tracker for Generic Object Tracking

基于匹配的跟踪算法能够将特定目标的识别问题转化为模板匹配问题,具有较高的响应速度和跟踪精度,这使它在通用目标跟踪中占有优势.然而,此类算法缺乏在线适应性和对特定数据的针对性,难以应对目标和跟踪场景的复杂变化.针对这一问题,提出一种基于图结构的图记忆跟踪器以提升通用目标跟踪的准确性.首先,利用图的节点匹配机制实现目标先验知识与搜索输入的点对点局部匹配,并根据匹配结果定位目标位置.其次,利用目标位置信息生成新模板.为抑制相似实例的干扰,根据相似实例分类响应呈现多峰的特点对新模板进行动态筛选.最后,将经过筛选的新模板作为候选信息存入存储模块.为了防止筛选失误引起后续错误叠加、减少错误信息的参与度,存储模块对候选信息进行实时更新.视频序列上的测试结果显示,图记忆跟踪器的存储模块能够及时更新候选信息,保留目标不同时刻的状态.常用跟踪基准上的结果显示,图记忆跟踪器在成功率和精度上优于基于匹配的基线跟踪器SiamRPN.与最近的先进跟踪器CstNet相比,图记忆跟踪器在OTB100基准上获得了11.75%的重叠成功率增益,10.53%的精度增益,在VOT2016基准上获得了8.59%的预期平均重叠增益.速度测试的结果显示,图记忆跟踪器能够在单片RTX2070上实现29帧/s的运行速度.

Matching-based tracking algorithm transforms the issue of detecting specific targets into template matching and has a high response speed and tracking precision.As a result,it has an advantage in generic target tracking.However,lack of online adaptability,applicability to specific data,and difficulty in coping with the complex changes in targets and tracking scenes are some of the issues it faces.To address this issue,a graph structure-based graph memory tracker is presented to enhance the accuracy of generic object tracking.First,the node-matching mechanism of the graph was used to achieve point-to-point local matching between the prior knowledge of the target and the search input.The target position was then located using the matching result.Second,the new template was created by processing the target location information.To suppress the interference from similar objects,the new template was dynamically screened according to the characteristics of the multi-peak classification response of similar objects.Finally,the new screened template was saved as candidate information in the memory module.The memory module updated candidate information in real-time to prevent subsequent error superposition produced by screening errors and to decrease the participation of error information.The results from the video sequences demonstrated that the memory module of the graph memory tracker can update the candidate information in time and store the target’s state at each instance.The results on common tracking benchmarks reveal that the graph memory tracker is superior to the matching-based baseline tracker SiamRPN in success rate and accuracy.Compared with the latest advanced tracker CstNet,we achieved an 11.75%overlap success rate gain,10.53%accuracy gain on OTB100,and 8.59%expected average overlap gain on VOT2016.The graph memory tracker achieves a running speed of 29 frames per second on a single RTX2070 in the speed test.