CN104021564A - Adaptive mean shift algorithm based on local invariant feature detection - Google Patents

Abstract

The invention discloses an adaptive mean shift algorithm based on local invariant feature detection. According to the adaptive mean shift algorithm based on local invariant feature detection, local invariant feature detection and the adaptive mean shift algorithm are combined, the detection and matching of local invariant feature points of an object are introduced during searching, and the region of search is recalculated through obtained matched feature points, so that the region of search can be excellently constrained around a target range, and finally, the accuracy of a tracking process is ensured. The adaptive mean shift algorithm based on local invariant feature detection has the advantage that the accuracy and stability of searching are greatly improved relative to those of adaptive mean shift algorithms.

Description

A kind of self-adaptation mean shift algorithm detecting based on local invariant feature

Technical field

The present invention relates to the real-time tracing of moving object, specifically a kind of method for tracking target self-adaptation mean shift algorithm used.

Background technology

Average drifting refers to selects certain point as starting point in image, calculate the mean difference that this puts current side-play amount, and this point is moved to the reposition that its side-play amount is pointed to, the position of usining after movement is as new starting point, continue to calculate and move, until arrive the position that meets constraint condition.Therefore this algorithm obtains optimum solution by iteration constantly.Mean shift algorithm is mainly for object, to carry out a kind of method for tracking target of non-rigid motion.This method is described target area with a kind of non-parametric histogram Density Estimator function, using Bhattacharyya coefficient as similarity criteria, constantly its mean shift vector is carried out to iteration, find the candidate region the highest with target area similarity, thereby realize the tracking to moving target.

Self-adaptation mean shift algorithm is the color probability distribution based on continuous dynamic change in image mainly, has good stability and real-time.It determines current location and the size of target in video image by color of object feature, the position getting and size are applied in the image of next frame, and the search window of initialization next frame, repeats this search procedure and can realize the tracking to target.Cardinal principle is exactly the variation along with the time, and when target moves, color probability distribution also can change along with the variation of time, by the change calculations to color of image probability distribution, realizes the real-time follow-up to target.

Self-adaptation mean shift algorithm is a kind of nonparametric technique of rising based on density gradient, by interative computation, finds target location, and realize target is followed the tracks of.Its significant advantage is that algorithm calculated amount is little, is simple and easy to realize, and is well suited for real-time follow-up occasion; But usually failure while following the tracks of little target and Fast Moving Object, and all under circumstance of occlusion, can not self-recovery follow the tracks of.When color and the target of background close, or target proximity has relatively object of the algorithm close with the tone of target, and search window can be included automatically, causes tracking window to expand, even sometimes tracking window is expanded to whole video frame, finally cause following the tracks of unsuccessfully.

Summary of the invention

The object of the invention is to be the deficiency for above-mentioned prior art, a kind of self-adaptation mean shift algorithm detecting based on local invariant feature is proposed, with local invariant feature, detect to retrain the search window of self-adaptation mean shift algorithm, reach accurate target following.

The self-adaptation mean shift algorithm detecting based on local invariant feature, is characterized in that, comprises the following steps:

Step 1: at the start frame of video the target area that selection will be followed the trail of ;

Step 2: by image from RGB color space, be mapped to hsv color space, and calculate target area color probability distribution;

Step 3: calculate and extract target area local invariant feature point;

Step 4: calculate respectively according to probability distribution zeroth order square with first moment , first moment , calculate the barycenter of search window , ;

Step 5: at next frame in with as the starting point of search window, with for search window, search window size meet , get the odd number of approximation;

Step 6: calculate the local invariant feature point in region of search ;

Step 7: relatively with obtain the unique point of coupling , and rebuild according to the unique point of coupling target area in frame ;

Step 8: return to step 4, carry out repeat search, meet search stop condition until follow the tracks of end person.

 

Further, described step 3 comprises the following steps:

First build metric space wherein for scale factor, * is convolution algorithm symbol, for Gaussian function, calculate difference of Gaussian space, ;

Secondly differentiate between difference empty is drawn to extreme point , wherein it is the skew with respect to sampled point; Passing threshold , with Hessian matrix , remove low contrast point and marginal point; According to the gradient magnitude of pixel in neighborhood and deflection determine the direction of unique point;

Finally choose neighborhood window in subregion, the vector information of 8 directions in every sub regions is sorted successively, form one the proper vector of dimension, i.e. SIFT feature descriptor; .

 

Beneficial effect of the present invention: self-adaptation mean algorithm is in search procedure, because constantly expanding or block, search window can cause error, so the present invention introduces local invariant feature and detects in self-adaptation mean shift algorithm, thereby improve search precision, reach accurate target following.

Accompanying drawing explanation

The self-adaptation mean shift algorithm search routine figure that Fig. 1 detects based on local invariant feature.

Specific implementation method

As shown in Figure 1, the invention provides a kind of self-adaptation average drifting method detecting based on local invariant feature, the concrete implementation step of the method is as follows:

Step 1: at the start frame of video the target area that selection will be followed the trail of .

Step 2: by image from RGB color space, be mapped to hsv color space, and calculate target area color probability distribution.

Step 3: calculate and extract target area local invariant feature point.

First build metric space wherein for scale factor, * is convolution algorithm symbol. for Gaussian function.Calculate difference of Gaussian space ;

Secondly differentiate between difference empty is drawn to extreme point , wherein it is the skew with respect to sampled point; Passing threshold , with Hessian matrix , remove low contrast point and marginal point; According to the gradient magnitude of pixel in neighborhood and deflection determine the direction of unique point;

Finally choose neighborhood window in subregion, the vector information of 8 directions in every sub regions is sorted successively, form one the proper vector of dimension, i.e. SIFT feature descriptor; .

Step 4: calculate respectively according to probability distribution zeroth order square with first moment , first moment .Calculate the barycenter of search window , .

Step 5: at next frame in with as the starting point of search window, with for search window.Search window size meet , get the odd number of approximation.

Step 6: the local invariant feature of region of search point in calculating .

Step 7: relatively with obtain the unique point of coupling , and rebuild according to the unique point of coupling target area in frame .

Step 8: return to step 4, carry out repeat search, finish or meet to search for stop condition until follow the tracks of.

Claims (2)

1. the self-adaptation mean shift algorithm detecting based on local invariant feature, is characterized in that, comprises the following steps:

Step 1: at the start frame of video the target area that selection will be followed the trail of ;

Step 2: by image from RGB color space, be mapped to hsv color space, and calculate target area color probability distribution;

Step 3: calculate and extract target area local invariant feature point;

Step 4: calculate respectively according to probability distribution zeroth order square with first moment , first moment , calculate the barycenter of search window , ;

Step 5: at next frame in with as the starting point of search window, with for search window, search window size meet , get the odd number of approximation;

Step 6: calculate the local invariant feature point in region of search ;

Step 7: relatively with obtain the unique point of coupling , and rebuild according to the unique point of coupling target area in frame ;

Step 8: return to step 4, carry out repeat search, meet search stop condition until follow the tracks of end person.

2. a kind of self-adaptation mean shift algorithm detecting based on local invariant feature according to claim 1, is characterized in that: described step 3 comprises the following steps:

First build metric space wherein for scale factor, * is convolution algorithm symbol, for Gaussian function, calculate difference of Gaussian space ;

Secondly differentiate between difference empty is drawn to extreme point , wherein it is the skew with respect to sampled point; Passing threshold , with Hessian matrix , remove low contrast point and marginal point; According to the gradient magnitude of pixel in neighborhood and deflection determine the direction of unique point;

Finally choose neighborhood window in subregion, the vector information of 8 directions in every sub regions is sorted successively, form one the proper vector of dimension, i.e. SIFT feature descriptor; .