采用RGB-D时空上下文模型的多目标遮挡跟踪算算法

Multi-target occlusion tracking algorithm employing RGB-D spatio-temporal context model

为了提高实时RGB-D目标遮挡跟踪精确度,解决多目标遮挡跟踪容易发生模型漂移和跟踪丢失等问题,本文提出一种基于RGB-D时空上下文模型的多目标遮挡跟踪算法.首先获取多目标检测定位区域,再通过目标时空上下文特征提取,建立目标时间上下文模型、目标空间上下文模型构成目标RGB-D时空上下文模型;然后在跟踪器判别跟踪状态时通过计算时间一致性进行颜色和深度特征自适应融合确定目标在当前帧位置;最后,当跟踪器判别多目标遮挡时引入深度概率,利用深度概率信息特征进行约束,通过最大后验概率(MAP)关联模型有效解决目标遮挡跟踪问题.在公用数据集clothing store dataset和princeton tracking benchmark dataset上进行定性对比实验和定量结果分析表明,本文提出的算法具有良好的遮挡跟踪性能,能较好解决多目标遮挡跟踪问题,提高目标遮挡跟踪的精确性和鲁棒性.

In order to improve the accuracy of real-time RGB-D target occlusion tracking and solve the problems of model drift and tracking loss in multi-target occlusion tracking,this paper proposes a multi-target occlusion tracking algorithm based on RGB-D spatio-temporal context model.Firstly,the multi-target detection and location region is obtained,and then the target temporal context model and the target spatial context model are used to establish the target RGB-D spatio-temporal context model through target spatio-temporal context feature extraction.Then,when the tracker judges the tracking state,the color and depth features are adaptively fused by calculating the time consistency to determine the target position in the current frame.Finally,when the tracker discriminates multi-target occlusion,the depth probability is introduced,and the depth probability information feature is used to constrain.The maximum a posteriori(MAP)correlation model is used to effectively solve the problem of target occlusion tracking.The qualitative comparison experiments and quantitative results on public datasets clothing store dataset and Princeton tracking benchmark dataset show that the proposed algorithm has good occlusion tracking performance,can better solve the problem of multi-target occlusion tracking,and improve the accuracy and robustness of target occlusion tracking.