基于改进UpdateNet的行人视觉跟踪算法

Pedestrian Visual Tracking Algorithm Based on Improved UpdateNet

目前,基于孪生网络的跟踪器主要是把跟踪过程看作是目标模板分支和待检测区域分支的一种互相关计算,它能够在速度和精度方面达到一个良好的均衡。然而,在跟踪过程中当前帧模板是对前一帧累积模板的线性结合,目标遮挡的问题一直很难被解决。针对该问题,本文采用改进SiamRPN跟踪器并融合UpdateNet网络的方案实现对行人单目标的跟踪。首先利用改进SiamRPN网络模块生成线性模板,然后在此基础上融合UpdateNet网络生成更新模板并进行多阶段训练,最后根据数据集的评价指标选取最优参数模型完成行人跟踪任务。本文在基准数据集OTB2015和其子集上进行实验,实验结果表明,本文所采用的方法取得的效果相比原来方法有明显提升,精度和成功率分别提高2.1和1.6个百分点,而且保持了实时跟踪帧率,同时在解决遮挡问题方面超过了DaSiamRPN、SiamDW等先进方法.

At present,trackers based on siamese networks mainly regard tracking as a cross-correlation calculation between the target template branch and the branch of the region to be detected,which can achieve a good balance between speed and accuracy.However,in the tracking process,the current frame template is a linear combination of the previous cumulative frame template,which leads to the object occlusion difficult to solve.In order to cope with this thorny problem,we adopt the improved SiamRPN tracker and integrate UpdateNet network to track the single pedestrian target.Firstly,the improved SiamRPN network module is used to generate the linear template,then the UpdateNet network is integrated to generate the updated template and perform multi-stage training.Finally,the optimal parameter model is selected to complete the pedestrian tracking task,according to the evaluation index of the dataset.We make the experiment in the benchmark data sets of the OTB2015 and its subset,the results show that the proposed method has obvious improvement than the original method,accuracy and success rate are increased by 2.1 and 1.6 percentage points respectively,while the real-time tracking frame rate is kept.It is also better than many advanced methods to deal with occlusion,such as DaSiamRPN,SiamDW,etc.