基于两级隐式形状模型的抗遮挡目标跟踪

摘要

为解决遮挡条件下的目标跟踪问题，提高定位精度，提出一种基于两级隐式形状模型的目标跟踪算法。利用 Fast Hessian 检测子提取待跟踪目标及周围区域的局部关键点区域构建码本字典，用 SURF 描述子获取码本的特征描述矢量建立码本支持模型，利用广义 Hough 变换建立码本字典集与目标之间的共生关系，通过隐式形状模型进行在线学习更新，通过寻找投票空间中的极大值来对目标进行定位，根据跟踪过程中目标遮挡程度的不同，分别赋予目标自身码本投票及周围码本投票不同权重，提高不同遮挡状态下目标定位精度。实验结果表明，在目标被遮挡甚至不可见或者丢失后重新回到视场时，该算法均能鲁棒地定位出目标。%A target tracking algorithm based on two-level implicit shape model is proposed to solve the tracking problem under occlusion and improve the location accuracy.Firstly,partial key points about the target and surrounding areas were extracted to build the codebook dictionary by Fast Hessian detector,and the feature description vectors of the codebook were constructed by SURF descriptor to establish the codebook support model.Secondly,the symbiotic relationship between the codebook dictionary set and target was established through the generalized Hough transform,and the online updating was accomplished by the implicit shape model.Finally,by finding the maximum value in the voting space,the target was located.According to the occlusion states in the tracking process,different voting weights were assigned to the codebook of target itself and surrounding area respectively,in order to improve the location accuracy under different occlusions.Experiments show the algorithm can locate the target robustly even though the target is occlusive,or even not visible,or returns to the field of view after missing.