一种改进的两阶段目标跟踪方法

An Improved Two-phase Object Tracking Method

针对动态场景中使用固定模板进行跟踪容易丢失目标的问题以及利用动态模型估计目标位置时产生的漂移问题,提出一种改进的基于偏最小二乘法的两阶段目标跟踪方法。该方法利用偏最小二乘分析法对在高维特征空间中搜集的正负样本降维,获得特征子空间构建目标表观模型集。跟踪在贝叶斯推理框架下进行:在初始跟踪阶段,利用粒子滤波原理及似然函数估计目标的初步位置;在校正阶段,采用一种适应性的基准模型确定最终的目标位置。对一些视频序列的实验结果证明了所提出方法的有效性。

To solve the problem that tracking with fixed templates is prone to fail in dynam- ic scenes and alleviate the visual drift problem caused by using dynamic models to estimate the target position in object tracking, an improved two-phase object tracking is proposed by partial least squares method. A low-dimensional feature subspace is studied with a few posi- tive and negative samplcs in the high-dimensional feature space via partial least squares （ PLS ） analysis, which constructs an appearance model. Object tracking is carried out within the Bayesian inference framework：in the initial tracking phase, the preliminary estimation of object location is achieved by particle filter principle;in calibrating phase, the adaptive benchmark model is adopted to determine the final target location. Experimental results on some video sequences show the proposed method effectiveness.