结合标准对冲与核函数稀疏分类的目标跟踪

针对经典稀疏分类目标跟踪算法在噪声,遮挡等恶劣环境下精度不高的问题,提出了一种新的目标跟踪算法。该算法在标准对冲框架下结合了核函数稀疏分类方法以及自适应字典更新方法,能够较好地适应类间相似度较高与目标外形变化较大等恶劣情况。核函数技巧能够增强分类器性能,但通用方法求解凸优化问题的效率较低,不能满足目标跟踪问题的实时性要求,故提出用核函数随机坐标下降(KRCD)算法来高效求解稀疏系数,并使用核函数稀疏分类方法(KRCD-SRC)来计算各个粒子的代价值。为了避免模板漂移问题,解释了目标字典和背景字典的在线更新方法。最后,结合标准对冲算法估算目标的状态信息。在使用50个粒子进行跟踪时,本文算法的处理帧率能够达到14frame/s。相比其它几种经典目标跟踪算法,本文算法具有更好的精确性和鲁棒性。

Visual object tracking combined normal hedge and kernel sparse representation classification

To achieve the robust tracking for a visual object under challenging conditions in the noisy, occlusion and the deformation, a novel visual object tracking method is proposed in this paper. By combining the Kernel Sparse Representation Classification (KSRC) and adaptive dictionary updating method under Normal Hedge framework, this method can handle tough situations like high inter-class similarities and drastically target appearance variations. Although the KSRC enhances classification performance, standard convex optimization is not fast enough for tracking in real time. Thus an efficient Kernel Random Coordinate Descent(KRCD) method is proposed to calculate the sparse coefficient vector, and the KRCD-SRC classification method is taken to calculate the loss value of each particle. In order to avoid the template drifting, the adaptive dictionary updating method is also given. At last, the states of the target are estimated by the Normal Hedge. Experiments show that the average computing frame rate of the proposed method is 14 frame/s when 50 particles are used. Extensive test results suggest that the proposed method outperforms several state-of-art tracking methods in many complex conditions.