一种基于线性判别空间增量更新的跟踪方法  

Object Tracking Method Based on Incrementally Updating Linear Discriminant Subspace

作　　者：钱诚[1] 徐舒畅[1] 张引[1] 张三元[1] 

机构地区：[1]浙江大学计算机科学与技术学院,杭州310027

出　　处：《模式识别与人工智能》2010年第2期250-255,共6页Pattern Recognition and Artificial Intelligence

基　　金：国家863计划资助项目(No.2007AA01Z311;2007AA04Z1A5)

摘　　要：为了在视频中稳定地跟踪目标物,提出一种基于增量型线性判别分析的目标跟踪方法.该方法利用一组仿射参数描述目标物在视频中的空间位置及姿态,根据状态转移模型预测得到下一帧视频中目标物的候选图像样本集合.计算各样本在线性判别空间中为目标图像的似然度,以具有最大似然度的样本作为目标图像区域.最后由类间散度矩阵与类内散度矩阵的充分生成集作旋转变换完成投影矩阵的增量更新,以保持判别空间的判别能力.实验结果表明,该方法对目标物及其周围背景的外观变化具有较强的自适应性,能够有效地对运动目标进行仿射不变的跟踪.A method for object tracking is presented to track objects steadily based on incremental linear discriminant analysis. The locations and poses of the object are represented by a set of affine parameters. Resorting to a state transition model, a group of image samples are predicted as candidates of the image patches of the object in the next frame. Their likelihoods of being the object image patch are measured after they are projected into a linear discriminant subspaee. Then a sample with maximum likelihood is regarded as the＇ object image patch. Finally, sufficient spanning sets of total scatter matrix and between-class scatter matrix are rotated to update projection matrix for maintaining the discrimination power of the subspaee. Experimental results show that the method is robust to variation in appearances of objects and surrounding background, and it is available in affine invariant tracking.

关 键 词：目标跟踪 增量型线性判别分析 生成集 仿射不变 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]