基于核主元投影的视觉跟踪算法

Visual Tracking Based on Kernel Principal Component Projection

视觉跟踪优化问题,在应用领域得到广泛研究。为了适应视觉场景中目标表观的无规律变化,提出了一种核主元投影的视觉跟踪算法。在初始跟踪阶段,采用粒子滤波相关跟踪算法获取场景中目标的先验表观图像数据,根据目标表观图像颜色与梯度方向联合直方图定义核函数,计算核矩阵,利用核主元分析得到图像样本在核空间的投影矩阵,最后通过贝叶斯滤波获得目标状态的最佳估计,并进行对比实验。实验结果说明了改进算法在真实的视频监控场景且遮挡环境下具有鲁棒性,对目标尺度变化具有自适应性。

Visual tracking is the research focus in computer vision,with widespread applications.In order to adapt the irregular changes of object apparent in visual scene,we proposed a visual tracking algorithm based on kernel principal component projection.In the initial stages of tracking,the particle filter algorithm was used to get the object apparent image.And based on the joint histograms of color and oriented gradients,the kernel function was defined and the kernel matrix was calculated.Then kernel principal component analysis was used to obtain the projection martix of image samples in kernel sapce.Finally,the best estimation of the object state was obtained through bayesian filter and simulation.The experimental results show that the proposed algorithm is of robustness in real video surveillance scene and occlusion environment,meanwhile is adaptive to object scale changes.