CN110232704A - Dimension self-adaption anti-shelter target tracking based on optimal characteristics - Google Patents

Abstract

Scale-adaptive anti-occlusion object tracking method based on optimal features. Traditional compressed tracking methods fail to solve the problems of optimal feature extraction, scale adaptation and occlusion judgment. The method of the present invention constructs the optimal feature pool for the compressed features by introducing the genetic algorithm, sorts and selects the optimal features, extracts the SURF feature to construct the reference point model and the target feature library, updates the window scale through affine transformation, and simultaneously sets The warning window judges whether it is blocked, uses the reference point model to estimate the target position for the blocked situation, and finally completes the update according to the tracking result until the tracking ends. The invention can improve the accuracy of the traditional compressed tracking method and solve the problems of scale adaptation and occlusion.

Description

Scale Adaptive Anti-Occlusion Target Tracking Method Based on Optimal Features

Technical field:

The invention belongs to the field of computer vision, and specifically relates to a moving target tracking method.

Background technique:

Target tracking technology is one of the important technologies of intelligent video surveillance, and it has important research significance and value no matter in military defense or people's livelihood security issues. In military defense, video-based human-computer interaction and intelligent visual navigation technology have always been the focus of researchers. In terms of people's livelihood and safety, monitoring systems in public places such as residential areas provide reliable protection for the safety of people's lives and properties. The intelligent transportation system on the street guarantees the smooth traffic and the safety of people's travel, and provides all-round protection for the people.

The target tracking technology based on online learning trains and updates the classifier while tracking, which can better adapt to the change of the target. The advantages of online learning tracking technology are obvious, making it quickly become the focus of research in the field of tracking. With the advancement of science and technology and the interdisciplinary applications between disciplines are becoming more and more common, the principle of compressed sensing in the field of signal processing has been introduced into the field of object tracking in recent years. The redundant information of image extraction features can be reduced through compressive sensing theory, and image features can be compressed, which provides a basis for implementing online learning algorithms to update classifiers. The compressed tracking algorithm proposed by Kaihua Zhang et al. directly applies the compressed matrix after the sparse matrix projection to classify the target. It has obvious advantages, but there are also some shortcomings.

Invention content:

The purpose of the present invention is to provide a scale-adaptive anti-occlusion target tracking method based on optimal features to solve the problems of insufficient feature effectiveness, fixed target frame size and inability to judge occlusion in existing compression tracking methods.

The purpose of the present invention can be achieved through the following technical solutions:

A scale-adaptive anti-occlusion object tracking method based on optimal features, said tracking method comprising the following steps:

Step 1: Calibrate the target area on the initial frame, generate a series of positive and negative samples to calculate the probability distribution of compressed feature training, extract SURF features and randomly select a fixed number of feature points as the target feature library and construct a reference point model;

Step 2: Input the t-th frame image, find all possible target frames within a certain search radius according to the results of the previous frame and extract compressed features, and construct the optimal feature pool by genetic algorithm for sorting;

Step 3: Select the first n groups of features from the feature pool and send them to the Naive Bayesian classifier to calculate the response value, take the maximum value as the tracking result of the current frame, extract SURF feature points to match with the target feature library, and use the matching feature points to solve the target affine Transform parameters and update the tracking window;

Step 4: Expand the target window by 8 pixels to set the warning window and divide it into two symmetrical parts. Calculate the color histogram of the two parts and perform normalization processing to determine whether it is blocked. If the blocking condition is met, use the SURF reference point model to infer The target position is updated with the final tracking result, and the above steps are repeated until the end of the video.

Further, the process of step one is as follows:

The target area described in step 1 is manually calibrated at the initial frame, and the generated positive and negative samples use V=PX compressed features, P represents the sparse matrix of compressed feature dimensions, V and X represent high-dimensional and low-dimensional compressed features, respectively, P( v=1), P(y=0) represent the prior probability of the target or sample respectively, and the conditional probability distribution is defined as and Represents the mean and standard deviation of the features calculated in the target sample according to the normal distribution, and Representative features Calculate the mean and standard deviation of the feature values conforming to the normal distribution in the background sample, extract SURF features, randomly select 4 feature points as the target feature library, and randomly select feature points at _a certain distance from the target center as the reference point set DB and Voting set D _V .

Further, the process of the second step is as follows:

Input the t-th frame image, use D ^γ ={Z||l(z)-l _t-1 ||<γ} to search for the target candidate frame, where γ is the search radius, l _t-1 is the target position in the previous frame, The feature pool is constructed and sorted by genetic algorithm, and the specific steps are as follows:

Step1: Randomly generate N sets of compressed features, and the number of features used as a classifier each time is n;

Step2: Introduce a selection operator to sort the effectiveness of features, select the first n groups as classifiers and directly retain them for the next generation, and clear the other N-n groups of features;

Step3: Introduce a crossover operator to randomly select 2s excellent individuals for crossover, and generate 2s crossover individuals to be included in the feature pool;

Step4: Introduce a mutation operator to select m excellent individuals to randomly generate mutations and include them in the feature pool;

Step5: Introduce the migration operator to re-randomly generate mi individuals into the feature pool;

use the formula

For normal distribution distance sorting, the larger the distance, the better the validity.

Further, the process of step three is as follows:

The input of n sets of excellent features adopts Calculate the response value, P(v _i |y=1) and P(v _i |y=0) represent the conditional probability of positive and negative classes respectively, extract SURF feature points and match the target feature library to get 4 sets of matching features, using the formula

I _t =H×I _t-1

Calculate the affine transformation parameters of two adjacent frames to update the tracking window, H is the affine transformation matrix, ρ is the scale factor, and θ is the rotation angle.

Further, the process of step four is as follows:

Expand 8 pixels outside the target frame to set a slightly larger warning window, which is divided into left and right parts. The left and right tracking windows are respectively expressed as T-W(k), k=1, 2; the left and right warning windows are respectively expressed as A-W(k), k= 1, 2, 1 represent the left half, 2 represents the right half, calculate the color histogram of the two parts at time t and normalize it, use the Bhattacharyya coefficient to measure the similarity of the color histogram to judge the occlusion, the formula is as follows:

If three consecutive frames satisfy ρ ₂ (k, t) > ρ ₁ (k, t), mark it as True to consider occlusion, and False to leave occlusion and satisfy The feature points are included in the reference set, satisfying Include the voting set, d is the SURF descriptor of the feature f, vote the feature points satisfying the reference set according to P(X|I)∞S=∑ _f∈F P(X|f)P(X|I) and accumulate the results , the maximum point is the inferred position, search for the required positive samples within a certain radius of the tracking results, and extract negative samples away from the target area, according to the formula

Probability distribution updates for target and background regions are done.

The beneficial effects of the present invention are:

1. Introduce the optimal search strategy of genetic algorithm, simulate the evolution process of organisms through selection, crossover, mutation and migration operators, express the adaptability of each individual to the current environment through the fitness function, and construct the optimal characteristics of the target characteristics The pool is updated in real time to ensure the validity of the features and solve the problem of poor stability of the classifier caused by unstable feature selection.

2. Extract the SURF feature of the target to construct the target feature library, update the size of the tracking frame through the affine transformation between adjacent frames, and solve the problem of incorrect sample extraction when the target changes caused by the fixed tracking frame.

3. Add an occlusion judgment mechanism, and improve the effectiveness of tracking by setting a warning window to detect the occurrence of occlusion and whether to leave the occlusion.

4. Establish a reference point model with a certain fixed relationship with the target through the extracted SURF features. When occlusion occurs, the reference point is used to infer the target position, which improves the robustness under occlusion.

Description of drawings:

Fig. 1 is a flowchart of the present invention;

Fig. 2 is a schematic diagram of an occlusion judgment warning window.

Detailed ways:

Below in conjunction with accompanying drawing, the present invention will be further described.

As shown in Figure 1-2, the scale-adaptive anti-occlusion target tracking method based on optimal features, the specific steps of the method are as follows:

Step 1: Calibrate the target area on the initial frame, generate a series of positive and negative samples to calculate the probability distribution of compressed feature training, extract SURF features, randomly select a fixed number of feature points as the target feature library, and construct a reference point model.

In a specific embodiment, step one includes:

The target area described in step 1 is manually calibrated at the initial frame, and the generated positive and negative samples use V=PX compressed features, P represents the sparse matrix of compressed feature dimensions, V and X represent high-dimensional and low-dimensional compressed features, respectively, P( y=1), P(y=0) represent the prior probability of the target or sample respectively, and the conditional probability distribution is defined as and Represents the mean and standard deviation of the features calculated in the target sample according to the normal distribution, and Representative features Calculate the mean and standard deviation of the feature values conforming to the normal distribution in the background sample, extract SURF features, randomly select 4 feature points as the target feature library, and randomly select feature points at _a certain distance from the target center as the reference point set DB and Voting set D _V .

Step 2: Input the t-th frame of image, find all possible target frames within a certain search radius according to the results of the previous frame and extract compressed features, and construct the optimal feature pool by genetic algorithm for sorting.

In a specific embodiment, step 2 includes:

Input the t-th frame image, use D ^γ = {Z丨||l(z)-l _t-1 ||<γ} to search for the target candidate frame, where γ is the search radius, and l _t-1 is the target position in the previous frame , using the genetic algorithm to construct and sort the feature pool, the specific steps are:

Step1: Randomly generate N sets of compressed features, and the number of features used as a classifier each time is n;

Step2: Introduce a selection operator to sort the effectiveness of features, select the first n groups as classifiers and directly retain them for the next generation, and clear the other N-n groups of features;

Step3: Introduce a crossover operator to randomly select 2s excellent individuals for crossover, and generate 2s crossover individuals to be included in the feature pool;

Step4: Introduce a mutation operator to select m excellent individuals to randomly generate mutations and include them in the feature pool;

Step5: Introduce the migration operator to re-randomly generate mi individuals into the feature pool;

use the formula

For normal distribution distance sorting, the larger the distance, the better the validity.

Step 3: Select the first n groups of features from the feature pool and send them to the Naive Bayesian classifier to calculate the response value, take the maximum value as the tracking result of the current frame, extract SURF feature points to match with the target feature library, and use the matching feature points to solve the target affine Transform parameters and update the tracking window.

In a specific embodiment, step three includes:

The input of n sets of excellent features adopts Calculate the response value, P(v _i |y=1) and P(v _i |y=0) represent the conditional probability of positive and negative classes respectively, extract SURF feature points and match the target feature library to get 4 sets of matching features, using the formula

I _t =H×I _t-1

Calculate the affine transformation parameters of two adjacent frames to update the tracking window, H is the affine transformation matrix, ρ is the scale factor, and θ is the rotation angle.

Step 4: Expand the target window by 8 pixels to set the warning window and divide it into two symmetrical parts. Calculate the color histogram of the two parts and perform normalization processing to determine whether it is blocked. If the blocking condition is met, use the SURF reference point model to infer The target position is updated with the final tracking result, and the above steps are repeated until the end of the video.

In a specific embodiment, step 4 includes:

Expand 8 pixels outside the target frame to set a slightly larger warning window, which is divided into left and right parts. The left and right tracking windows are respectively expressed as T-W(k), k=1, 2; the left and right warning windows are respectively expressed as A-W(k), k= 1, 2, 1 represent the left half, 2 represents the right half, calculate the color histogram of the two parts at time t and normalize it, use the Bhattacharyya coefficient to measure the similarity of the color histogram to judge the occlusion, the formula is as follows:

If three consecutive frames satisfy ρ ₂ (k, t) > ρ ₁ (k, t), mark it as True to consider occlusion, and False to leave occlusion and satisfy The feature points are included in the reference set, satisfying Include the voting set, d is the SURF descriptor of the feature f, vote the feature points satisfying the reference set according to P(X|I)∞S=∑ _f∈F P(X|f)P(X|I) and accumulate the results , the maximum point is the inferred position, search for the required positive samples within a certain radius of the tracking results, and extract negative samples away from the target area, according to the formula

Probability distribution updates for target and background regions are done.

The specific implementation of the present invention has been described in detail above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned implementation, within the knowledge of those of ordinary skill in the art, it can also be made without departing from the gist of the present invention. Variations.

Claims (5)

1. a kind of dimension self-adaption anti-shelter target tracking based on optimal characteristics, which comprises the following steps:

Step 1: to initial frame spotting region, a series of positive negative samples is generated and calculate compressive features training probability distribution, are mentioned It takes SURF feature to randomly select fixed number characteristic point as target feature library and constructs with reference to point model；

Step 2: input t frame image searches all possible target frames according to previous frame result around in certain search radius And compressive features are extracted, optimal characteristics pond is constructed by genetic algorithm and is ranked up；

Step 3: n group feature is sent into Naive Bayes Classifier and calculates response before selected characteristic pond, is maximized as present frame Tracking result is extracted SURF characteristic point and is matched with target feature library, solves target affine transformation using matching characteristic point and joins Number updates tracking window；

Step 4: 8 pixel setting warning windows are extended out with target window and are divided into symmetrical two parts, calculate two parts Color histogram is simultaneously normalized and judges whether to block, and speculates mesh with reference to point model using SURF if meeting obstruction conditions Cursor position is updated using final tracking result, repeats above step, until video terminates.

2. the dimension self-adaption anti-shelter target tracking according to claim 1 based on optimal characteristics, it is characterized in that: Target area described in step 1 is demarcated manually in start frame, and the positive negative sample of generation uses V=PX compressive features, and P indicates pressure The sparse matrix of contracting intrinsic dimensionality, V and X respectively indicate higher-dimension and low-dimensional compressive features, P (y=1), and P (y=0) respectively indicates mesh The prior probability of mark or sample, conditional probability distribution are defined as WithFeature is respectively represented in target Sample calculates the mean value and standard deviation that characteristic value meets normal distribution,WithIt represents feature and calculates characteristic value in background sample Meet the mean value and standard deviation of normal distribution, extracts SURF feature, randomly select 4 characteristic points as target feature library, at random The characteristic point of selected distance target's center certain distance collection D as a reference point_BAnd ballot collection D_V。

3. the dimension self-adaption anti-shelter target tracking according to claim 1 or 2 based on optimal characteristics, feature It is: input t frame image, using D^γ=Z | | l (z)-l_t-1| | < γ } search target candidate frame, wherein γ is search radius, l_t-1For previous frame target position, using genetic algorithm construction feature pond and sort, specific steps are as follows:

Step1: being randomly generated N group compressive features, and the Characteristic Number as classifier is n every time；

Step2: introducing selection operator and be ranked up to characteristic validity, selects preceding n group as classifier and directly remains into down A generation, other N-n group features are reset；

Step3: introducing crossover operator randomly selects 2s excellent individual and is intersected, and generates 2s intersection individual and is included in feature Pond；

Step4: introducing m excellent individual of mutation operator selection is randomly generated variation and is included in feature pool；

Step5: operator is migrated in introducing, and mi individual of generation is included in feature pool at random again；

Using formula

To normal distribution distance-taxis, the bigger validity of distance is better.

4. the dimension self-adaption anti-shelter target tracking according to claim 3 based on optimal characteristics, it is characterized in that: The outstanding feature of the n group of input usesCalculate response, P (v_i| y=1) and P (v_i|y =0) positive and negative class conditional probability is respectively indicated, SURF characteristic point is extracted and target signature storehouse matching obtains 4 groups of matching characteristics, use Formula

I_t=H × I_t-1

It calculates adjacent two frames affine transformation parameter and updates tracking window, H is affine transformation matrix, and ρ is scale factor, and θ is rotation Angle.

5. the dimension self-adaption anti-shelter target tracking according to claim 4 based on optimal characteristics, it is characterized in that: Expand 8 pixels one slightly larger warning windows of setting in target outer frame and be divided into left and right two parts, left and right track window is expressed as T-W (k), k=1,2；Left and right warning window is expressed as A-W (k), k=1,2, and 1 represents left-half, and 2 represent right half part, calculates T moment two parts color histogram and normalized are measured color histogram similarity using Bhattacharyya coefficient and are sentenced Disconnected to block, formula is as follows:

If continuous three frame meets ρ₂(k, t) > ρ₁(k, t) thinks to block labeled as Ture, and False then leaves and blocks, and meetsCharacteristic point is included in reference set, meets It is included in ballot collection, d is characterized the SURF descriptor of f, will meet the characteristic point of reference set by P (X | I) ∞ S=∑_f∈FP (X | f) P (X | I) simultaneously accumulation result of voting, it is computed position at maximum point, is searched in tracking result certain radius Required positive sample extracts negative sample far from target area, according to the following formula

The probability distribution for completing target and background region updates.