在线自适应选择子空间的红外目标跟踪方法

传统基于子空间的目标跟踪方法以能量大小为准则建立子空间，没有考虑目标与背景之间的鉴别性，当两者间存在近似外观分布时将降低跟踪系统的性能。考虑到红外图像信噪比、对比度不高等特点，提出了一种以评估目标与背景间可区分能力为基础的子空间选择方法，并将该方法有效嵌入到粒子滤波跟踪框架下实现对红外目标的鲁棒跟踪。首先利用采样粒子分布以及当前目标状态，综合衡量粒子与目标间的特征分布差异和粒子逼近目标的程度来评估不同子空间的鉴别性，然后选择鉴别性最优的子空间作为下帧的跟踪子空间，从而实现对红外目标进行子空间自适应选择的鲁棒跟踪。对多个复杂场景下的目标跟踪实验表明所提出的算法要优于传统基于增量子空间学习的跟踪算法。

IR object tracking method via online adaptive subspace selection

The subspace constructing strategy of classic subspace-based tracking schemes is to select appropriate subspaces with maximum energy, in this strategy the discriminability between the target and background is neglected, so when the target and background have similar appearance the tracking system's performance may be degenerated. To solve the problems of IR image's low SNR and low contrast, a novel subspace selecting method was proposed based on analyzing the discriminability between the target and background. The IR object tracking process was realized by the particle filter with the provided subspace selecting strategy. In this case, based on the prior knowledge of the particles distributions and the target state, different subspace's tracking ability by considering both the feature difference and the particles' approximation level to the target was estimated firstly, then the optimal subspaces were selected to realized the IR target tracking. Experiments on several complex scenes indicate that the proposed algorithm has better performance than the classic one.