基于L_1范数和最小软阈值均方的目标跟踪算法

基于传统稀疏表示的目标跟踪算法无法解决跟踪过程出现的遮挡及运动模糊等问题,提出一种基于L_1范数和最小软阈值均方的目标跟踪算法。首先用主成分分析(principal component analysis,PCA)基向量建模跟踪目标的表观变化,同时对表示系数进行L_1范数约束;其次对误差项采用最小软阈值方法进行显示求解,同时对观测模型的更新上考虑跟踪目标的遮挡因素;最后在贝叶斯框架下搭建目标跟踪算法。在14个具有挑战性的跟踪视频上的试验结果表明:与其他算法相比,本研究能够克服跟踪过程中遮挡、角度变化、尺度变化、光照变化等影响跟踪性能的因素,具有较高的平均覆盖率和较低的平均中心点误差。

Object tracking via L1 norm and least soft-threshold square

Due to the occlusion and motion blur in the traditional object tracking algorithm, a novel object tracking algorithm via L₁ norm and least soft-threshold square was proposed to solve the problem of the failure of object tracking based on sparse representation. Firstly, the appearances of the object were modeled by the PCA(Principal Component Analysis)basis vector and the representation coefficients were constrained by L₁ norm. Secondly, the trivial error was solved by the least soft-threshold square and the occlusion factor was taken account in the updation of the observation model. At last, the object tracking algorithm was developed in the Bayesian inference framework. Experiments were conducted on fourteen challenging videos and the experimental results showed that the proposed algorithm could cope well with the occlusion, angle variation, scale variation and illumination variation, with the higher average overlap rate and the lower average center point error, compared with the other tracking algorithm.