Overall Thinking 1. Use a random perception matrix that meets the compression perception rip conditions to perform dimensionality reduction for multi-scale images. 2. Then, use simple Naive Bayes to classify dimensionality reduction features.
Algorithm flow
1. At t frame, we sampled several images of the target (positive sample) and background (negative sample), and then performed multi-scale transformation on them, then, a sparse measurement matrix is used to reduce the dimensionality of the features of multi-scale images, and then the features after dimensionality reduction (including the target and background, which are binary classification problems) are used) train Naive Bayes classifier ().
2. At t + 1 frame, we sampled n scanning windows around the target position tracked in the previous frame (avoid scanning the entire image ), the sparse measurement matrix is used to reduce dimensions and extract features. Then, the naive Bayes classifier trained at the T frame is used for classification. The window with the largest classification score is considered as the target window. In this way, the target tracking from t frame to t + 1 frame is realized.
Relate workIssuse 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 set2 drift Problems
How to do1. Generate a random measurement matrix
Achlioptas [16] proved that this type of matrix withs = 2 or 3 satisfies the Johnson-lindenstrauss lemma. [17] It is proved that the restricted isometry property in compressive sensing of CS is met if JL conditions are met. in this way, we can reconstruct X from the V after dimensionality reduction, where V = RX and R are random matrices.
2. Obtain the X vector.
W and H are the width and length of the selected target. We use the above series of rectangles of different scales to generate a vector similar to Haar-like, which contains a total of WXH rectangle filters, convolution of each pixel to generate M = (Wh) ^ 2 X. Although M is very large, the random measurement matrix is very sparse and can be reduced to a very small n-dimensional value.
3. Perform dimensionality reduction as follows
4. Construct and update a classifier
Assume that the data after dimensionality reduction is an independent distribution, and use the following Naive Bayes for classification (4)
Since diaconis and Freedman [23] showed that the random projections of high dimen1_random vectors are almost always Gaussian, we assumed P (Vi | y = 1) andP (Vi | Y = 0) in the classifier to be Gaussian.
Parameters are updated incrementally using the following formula (6 ).
Histogram similar to Gaussian
5. The overall algorithm is
Discussion1 because the algorithm in this article is data-independent, it is not like generating models like 1-tracker [10] and compressive sensing tracker [9], and it does not need to store previous training samples; haar-like in the broad sense, unlike [9] [10] using holistic templates for sparse representation, the features in this article are more robust. 2 PCA and its variants are widely used in the generation of tracking models []. However, these methods are not robust to occlusion because they use holistic representation; and it may not be possible to update correctly with new observations; compression tracking does not have these problems in self-taught learning approaches, because the model using the random measurement matrix is data-independent; random projection is better than Principal Component analysis3 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 thetracking ambiguity problem
4 measurement matrix is data-independent and no noise is introduced by mis-aligned samples5 similar representations, e.g ., local Binary patterns [26] and generalized Haar-like features [8], have been shown to be more effective in handling occlusion.
Experiment1 used evaluation criteria 1 ROI 2 center location error2 algorithm combines the merits of generative (features ?) And discriminative (Bayes ?) Appearance models to account for scene changes
Real-time compressive tracking paper notes
