融合块显著质心描述和多级关联的多目标跟踪

Block-level saliency centroid representation and multi-level association based multi-target tracking

提出一种融合目标分块、显著质心建模和多级关联的多目标跟踪(multi-target tracking,MTT)方法,用于提高互遮挡、相似目标干扰场景中的跟踪鲁棒、准确性。利用自适应阈值背景差分检测运动区域;将目标区域分块,根据块中运动像素处背景差分值计算色彩显著度,建立运动、色彩显著质心模型;建立目标间、目标与运动检测间全局、块级数据关联,判别互遮挡目标及块,并据块遮挡矩阵更新目标模板;利用有效色彩和运动信息计算块质心转移向量及融合权值,获得目标全局质心转移向量以定位目标。实验结果表明该方法对互遮挡、相似目标干扰及外观变化的多目标均具有稳定跟踪性能。

A novel frame-work of multi-target tracking （MTT） based on block division, saliency centroid modeling and multiqevel association is presented to enhance the robustness and accuracy of tracker under occlu- sion among targets and disturbance caused by similar target. Selbadaptive threshold value based background difference is employed to detect motion regions. Based on block-division of target region, the block-level color saliency computed from background difference at each moving pixel, is utilized to model the centroid being with motion and color saliency. To discriminate the targets and blocks being in occlusion, the global and bolck associ- ations are established among tracked targets and motion regions, and the block occlusion matrix is built to up date the target model. The valid color and motion pixels are utilized to calculate the shifting vetor and fusion weight of each block centroid, then the global centroid shifting vetor is gained and used to locate the target. Ex- periments demonstrate that the proposed method is robust enough for tracking the multi-target in scenarios of occlusion, similar target disturbance and appearance change.