总体思想
1 利用符合压缩感知RIP条件的随机感知矩阵对多尺度图像进行降维
2 然后对降维的特征採用简单的朴素贝叶斯进行分类
算法主要流程
1 在t帧的时候,我们採样得到若干张目标(正样本)和背景(负样本)的图像片,然后对他们进行多尺度变换,再通过一个稀疏測量矩阵对多尺度图像特征进行降维,然后通过降维后的特征(包含目标和背景,属二分类问题)去训练朴素贝叶斯分类器()。
2 在t+1帧的时候,我们在上一帧跟踪到的目标位置的周围採样n个扫描窗体(避免去扫描整幅图像),通过相同的稀疏測量矩阵对其降维,提取特征,然后用第t帧训练好的朴素贝叶斯分类器进行分类,分类分数最大的窗体就觉得是目标窗体。
这样就实现了从t帧到t+1帧的目标跟踪。
Relate Work
Issuse of online tracking algorithms(update models with samples from observations in resent frames)
1 adaptive appearance models are data-dependent, but there does not exist sufficient amount of data for online algorithms at out set
2 drift problems
How To Do
1 生成随机測量矩阵
Achlioptas [16] proved that this type of matrix withs = 2 or 3 satisfies the Johnson-Lindenstrauss lemma. [17]证明,满足JL条件的话即满足CS的restricted isometry property in compressive sensing. 从而我们可以从降维后的v最好的重建x,当中v=Rx,R为随机矩阵
2 得到x向量
w,h是所选目标的宽长。我们用上面一系列不同尺度的矩形来对图像进行类似haar-like的向量生成,一共是wxh个rectangle filters,对每一个像素进行卷积,生成m=(wh)^2个x。尽管m非常大非常大。由于随机測量矩阵十分稀疏,能够降到非常小的n维
3 例如以下做降维处理
4 构建并更新分类器
如果降维后的数据是独立的分布,用下面朴素贝叶斯进行分类(4)
Since Diaconis and Freedman [23] showed that the random projections of high dimensional random vectors are almost always Gaussian。 we assumed p(vi|y=1)andp(vi|y= 0) in the classifier to be Gaussian.
參数採用下式(6)进行递增的更新
类似高斯的直观图
5 总体算法为
Discussion
1 由于本文算法是data-independent的,所以不像1-tracker [10] ,compressive sensing tracker [9] 这些生成模型,不须要存储曾经的训练样本;所採用广义的haar-like。不像[9][10]採用holistic templates for sparse representation,本文的特征更为鲁棒
2 PCA和它的变种广泛应用在了生成跟踪模型的方法里[1,6],但这些方法由于用的是holistic representation,对遮挡不鲁棒; 并且不一定能update correctly with new observations。压缩跟踪不存在self-taught learning approaches存在的这些问题,由于利用随机測量矩阵的这个模型是data-independent的;random projection
好于 principal component analysis
3 The tracking-by-detection methods often encounter the inherent ambiguity problems as shown in Figure below. Babenko et al. [8] introduced multiple instance learning schemes to alleviate the
tracking ambiguity problem
4 measurement matrix is data-independent and no noise is introduced by mis-aligned samples
5 Similar representations, e.g., local binary patterns [26] and generalized Haar-like features [8], have been shown to be more effective in handling occlusion.
Experiment
1 用到的评价标准 1 ROI 2 center location error
2 Algorithm combines the merits of generative(features?) and discriminative(bayes?) appearance models to account for scene changes