多模态目标的联合压缩跟踪

Joint Compressive Tracking for Multimodal Target

针对多模态目标跟踪中大多仅考虑单个图像的异种特征融合或不同模态图像的同种特征融合,为了使得这两者间能自然集成,提出基于联合压缩感知的多模态目标统一跟踪方法.通过将多模态跟踪问题转化为多重?2-范数不等式约束下的多?1-范数联合最小化问题,并设计了能求解该联合最小化问题的特定增广拉格朗日乘子算法,从而实现快速而精准的目标跟踪,可同时处理来自同一图像或不同模态图像中的多种不同特征间的融合,并能自由地添加或删除特征.此外,还提出了基于稀疏集中度指标的目标模板协同更新方案,来筛选出表现最优的目标模板.在DCU,OTCBVS,BEPMDS,OTB50和VOT-TIR等数据集上采用逐帧跟踪的方法进行实验,结果表明在跟踪精度、成功率以及速度3个评价指标上,文中方法的平均性能分别达到了0.96,0.91和3.48.

Multimodal tracking is a challenging task.Most of the existing methods only consider the fusion of different features from single image or the fusion of identical feature from different modal images.In order to integrate them naturally,a unified multimodal tracking framework based on joint compressive sensing is proposed in this paper.Our framework can integrate different features extracted from single image or different modal images,and provide the flexibility to arbitrarily add or remove feature.We formulate the multimodal tracking as a joint minimization problem of multiple?1-norms with inequality constraint of multiple?2-norms,and derive a customized augmented Lagrange multiplier algorithm to solve the minimization problem,so that efficient tracking with both low computational burden and high accuracy.Besides,a collaborative template update scheme induced by sparsity concentration index is developed to screen out the best templates throughout the tracking procedure.The experiments are carried out on DCU,OTCBVS,BEPMDS,OTB50 and VOT-TIR image datasets by frame tracking method,and the experimental results show the average tracking accuracy,success rate and speed of our method is 0.96,0.91 and 3.48 respectively.