CN103198493B - A kind ofly to merge and the method for tracking target of on-line study based on multiple features self-adaptation - Google Patents

A kind ofly to merge and the method for tracking target of on-line study based on multiple features self-adaptation Download PDF

Info

Publication number
CN103198493B
CN103198493B CN201310121576.9A CN201310121576A CN103198493B CN 103198493 B CN103198493 B CN 103198493B CN 201310121576 A CN201310121576 A CN 201310121576A CN 103198493 B CN103198493 B CN 103198493B
Authority
CN
China
Prior art keywords
feature
target area
target
new candidate
area
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Expired - Fee Related
Application number
CN201310121576.9A
Other languages
Chinese (zh)
Other versions
CN103198493A (en
Inventor
苏育挺
刘安安
刘晓伟
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Tianjin University
Original Assignee
Tianjin University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Tianjin University filed Critical Tianjin University
Priority to CN201310121576.9A priority Critical patent/CN103198493B/en
Publication of CN103198493A publication Critical patent/CN103198493A/en
Application granted granted Critical
Publication of CN103198493B publication Critical patent/CN103198493B/en
Expired - Fee Related legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Landscapes

  • Image Analysis (AREA)

Abstract

The invention discloses and a kind ofly to merge and the method for tracking target of on-line study based on multiple features self-adaptation, extraction target signature as template characteristic; Three kinds of features are extracted respectively to new candidate target region; Self-adaptation fusion is carried out according to the distinguishing property of each feature and correlativity; Calculate Pasteur's distance of feature and exemplary feature after merging, using after Pasteur's range normalization as the weight of new candidate target region; Carry out overlappingly judging to the new candidate target region of weight limit and target area, if Duplication is less than Duplication threshold value by many times of region input detectors of the new candidate target region of weight limit, when recognizer output represents and follows the tracks of successfully and upgrade recognizer, template characteristic and target area; If export no, represent and find fresh target; If Duplication is more than or equal to Duplication threshold value, upgrade recognizer, template characteristic and target area.Enhance the adaptive faculty of target following in different scene and certain deformation situation, avoid and block the rear easy problem that drift occurs to follow the tracks of.

Description

Target tracking method based on multi-feature adaptive fusion and online learning
Technical Field
The invention relates to the field of target tracking, in particular to a target tracking method based on multi-feature adaptive fusion and online learning.
Background
Video target tracking refers to the process of detection, characterization and trajectory extraction of targets in a video sequence. Video target tracking has practical application requirements in the fields of video monitoring, event analysis, human-computer interaction and the like.
Currently, the most advanced monitoring systems in the world cannot perfectly handle dynamic tracking tasks in complex scenarios, such as: deformation, occlusion, lighting changes, shadows, or tracking in crowded environments. Especially when partial occlusion occurs between targets and the targets are deformed, target tracking remains a challenge.
Tracking methods based on single features generally initialize a target region and extract arbitrary target features, for example: color features, search and match at the next frame. However, the method is difficult to process the tracking task under the complex background, and the target tracking and the subsequent track evaluation are not robust.
Therefore, a target tracking method based on multiple features is also provided in the prior art, and the method usually extracts features such as colors and edges and can better complete tracking tasks under certain complex conditions.
In the process of implementing the invention, the inventor finds that at least the following disadvantages and shortcomings exist in the prior art:
the target tracking method based on multiple characteristics cannot adapt to the change of the shape of the target and cannot well solve the problem of tracking drift after partial shielding occurs between the targets.
Disclosure of Invention
The invention provides a target tracking method based on multi-feature adaptive fusion and online learning, which avoids the problem of tracking drift, well adapts to the change of the appearance of a target and is described in detail in the following description:
a target tracking method based on multi-feature adaptive fusion and online learning comprises the following steps:
(1) selecting a target area from a frame of image, extracting target characteristics and using the target characteristics as template characteristics;
(2) the initialization recognizer inputs a next frame of image and initializes a candidate target area; acquiring a new candidate target area according to a transfer formula;
(3) respectively extracting three characteristics of color, edge and texture from the new candidate target area; performing self-adaptive fusion according to the discriminativity and the correlation of the features;
(4) calculating the Bhattacharyya distance between the fused features and the template features, and normalizing the Bhattacharyya distance to be used as the weight of the new candidate target area;
(5) sorting the N new candidate target areas according to the weight, if the resampling judgment value is larger than the resampling judgment threshold value, resampling, and executing the step (6); if not, executing the step (6);
(6) performing overlapping judgment on the new candidate target area with the maximum weight and the target area, and executing the step (7) if the overlapping rate is smaller than an overlapping rate threshold value; otherwise, executing step (8);
(7) inputting the multiple regions of the new candidate target region with the largest weight into a detector, and if the detector outputs 0, indicating that the tracking fails; otherwise, inputting the output result of the detector into a recognizer, and if the output of the recognizer is yes, indicating that the tracking is successful and updating the recognizer, the template features and the target area; if the output is not, indicating that a new target is found, and ending the process;
(8) and if the overlapping rate is more than or equal to the overlapping rate threshold value, the tracking is considered to be successful, the recognizer, the template characteristic and the target area are updated, and the process is ended.
The step of extracting the target feature and using the target feature as the template feature specifically comprises the following steps:
1) extracting color characteristic information;
2) extracting edge characteristic information;
3) extracting texture feature information;
4) and fusing the color feature information, the edge feature information and the texture feature information to obtain a target feature histogram as the template feature.
The step of extracting color feature information specifically includes:
1) dividing a color space into a colored area and an achromatic area, carrying out HSV (hue, saturation and value) partition on the colored area and the achromatic area to obtain QH×QSA color subinterval and QvAn achromatic subinterval, dividing said QH×QSA color subinterval and said QvOne achromatic subinterval as QH×QS+QvEach color interval u;
2) endowing different weights to the pixel points according to the distance between the pixel points and the central point of the target area, and voting the corresponding color interval u according to the HSV of the pixel points;
3) and counting the voting value of each color interval to obtain a color feature histogram.
The step of extracting the edge feature information specifically includes:
1) interpolating the target area to obtain an interpolation area 2 times as wide and as high as the target area, and then respectively blocking the target area and the interpolation area;
2) calculating the edge strength and the direction of each sub-block, dividing the edge direction into a plurality of direction areas within the range of 0-360 degrees, voting the direction areas according to the edge strength to obtain an edge feature histogram of each sub-block;
3) and connecting the edge feature histograms calculated by each sub-block to obtain a complete edge feature histogram.
The step of extracting the texture feature information specifically includes:
1) calculating a local binary pattern characteristic histogram for each sub-block;
2) and connecting the local binary pattern feature histograms calculated by each sub-block to obtain a complete texture feature histogram.
The discriminativity is defined as the similarity degree of the new candidate target area and the adjacent background with respect to a certain feature, and is represented by the Papanicolaou coefficient of two histograms.
The correlation is defined as the similarity degree of the new candidate target area and the template characteristic about a certain characteristic, and is represented by the Papanicolaou coefficient of two histograms.
The update identifier is specifically: the positive sample consists of a new candidate target area verified by the detector, and the negative sample is an area with the same size as the new candidate target area randomly selected from the background;
the updating template is characterized in that: taking the feature of the new candidate target region with the maximum weight as the updated template feature;
the update target area specifically includes: and taking the new candidate target area with the maximum weight as the updated target area.
The technical scheme provided by the invention has the beneficial effects that: the method overcomes the singleness of single characteristic, enhances the adaptability of target tracking under different scenes and certain deformation conditions, avoids the problem that tracking drift is easy to occur after shielding, and greatly improves the accuracy and robustness of target tracking.
Drawings
FIG. 1 is a flow chart of a target tracking method based on multi-feature adaptive fusion and online learning;
FIG. 2 is a schematic diagram of an initialization target area;
FIG. 3 is a schematic diagram of the occlusion of the object 1;
FIG. 4 is a schematic diagram of successfully retrieving an occlusion lost object through online learning;
FIG. 5 is a schematic diagram of another initialization target area;
FIG. 6 is a schematic diagram of the cross-occlusion of object 1 and object 2;
fig. 7 is a diagram illustrating that tracking drift does not occur in accurate tracking.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, embodiments of the present invention will be described in detail with reference to the accompanying drawings.
In order to avoid the problem of tracking drift and adapt to the change of the target shape well, the embodiment of the invention provides a target tracking method based on multi-feature adaptive fusion and online learning, the online learning can overcome the problems caused by target deformation and tracking drift through real-time automatic learning, and the expected tracking effect is realized well, see fig. 1, and the following description for details:
101: selecting a target area from a frame of image by taking a frame of any video sequence as input, extracting target characteristics and taking the target characteristics as template characteristics;
the operation of selecting the target area is well known to those skilled in the art, and the target area is a rectangular area, for example: according to actual requirements, a rectangular area where a tracked object is located can be manually selected; or the model of the detected object (for example, the human body detection model [1 ]) is adopted for automatic detection, and the template characteristic is calculated.
The steps of extracting the target feature and using the target feature as the template feature specifically comprise:
1) extracting color characteristic information;
the method adopts a kernel weighted color feature histogram based on an HSV (hue, saturation and brightness) color space model to model a target, and has the basic idea that:
(1) dividing the color space into a colored area and an achromatic area, performing HSV partitioning on the colored area and the achromatic area,obtaining QH×QSA color subinterval and QvAn achromatic subinterval, adding QH×QSA color subinterval and QvOne achromatic subinterval as QH×QS+QvEach color interval u;
for example: all brightness less than 20% or saturation less than 10% are included in the achromatic region and divided into Q values according to brightnessvAn achromatic sub-region, wherein the color regions other than the achromatic region are chromatic regions and are divided into Q regions according to chromaticity and saturationH×QSA color sub-interval.
(2) According to the distance between the pixel point and the central point of the target area, different weights are given to the pixel point (namely, a smaller weight is given to the pixel point far away from the target center, so that the interference of the target boundary and the background is weakened), and the corresponding color interval u is voted according to the HSV of the pixel point;
the target area is defined as a rectangular area with width W and height H. In order to increase the reliability of color distribution, the following function is adopted to distribute weight:
<math> <mrow> <mi>k</mi> <mrow> <mo>(</mo> <msub> <mi>x</mi> <mi>i</mi> </msub> <mo>)</mo> </mrow> <mo>=</mo> <mfenced open='{' close=''> <mtable> <mtr> <mtd> <mn>1</mn> <mo>-</mo> <msup> <mrow> <mo>(</mo> <mfrac> <mrow> <mi>y</mi> <mo>-</mo> <msub> <mi>x</mi> <mi>i</mi> </msub> </mrow> <mi>h</mi> </mfrac> <mo>)</mo> </mrow> <mn>2</mn> </msup> </mtd> <mtd> <mi>r</mi> <mo>&lt;</mo> <mn>1</mn> </mtd> </mtr> <mtr> <mtd> <mn>0</mn> </mtd> <mtd> <mi>r</mi> <mo>&GreaterEqual;</mo> <mn>1</mn> </mtd> </mtr> </mtable> </mfenced> </mrow> </math>
in the formula, k (x)i) Representing assigned pixel point xiY denotes the center point of the target area, xiThe ith pixel point in the target area is set;indicating the target area size.
The method determines the color interval corresponding to each pixel point through a Dirac function, (b (x)i) -u) represents a pixel point xiDistribution of color bins u in the histogram: b (x)i) Is a pixel point xiThe HSV of (1) is a compound of phi,eta is a constant. The dirac function is defined as follows:
(x-φ)=0,x≠φ
for example: there are 3 pixel points, do respectively: x is the number of1、x2And x3Pixel point x1The corresponding color interval is u1, the color interval corresponding to the pixel point x2 is u1, and the pixel point x is3If the corresponding color interval is u2, the voting result of the color interval u1 is the pixel point x1And pixel point x2The sum of the weights of (a); the voting result of the color interval u2 is a pixel point x3The weight of (c); the voting result of the color interval u3 is 0.
(3) And obtaining a color feature histogram through the voting value obtained by calculating each color interval, wherein the dimension of the color feature histogram is the sum of the numbers of the chromatic intervals and the achromatic intervals.
Color feature histogram pcCan be expressed as:
<math> <mrow> <msub> <mi>p</mi> <mi>c</mi> </msub> <mo>=</mo> <msub> <mi>C</mi> <mi>h</mi> </msub> <munderover> <mi>&Sigma;</mi> <mrow> <mi>u</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>U</mi> </munderover> <munderover> <mi>&Sigma;</mi> <mrow> <mi>i</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>N</mi> </munderover> <mi>k</mi> <mrow> <mo>(</mo> <msub> <mi>x</mi> <mi>i</mi> </msub> <mo>)</mo> </mrow> <mo>*</mo> <mi>&delta;</mi> <mrow> <mo>(</mo> <mi>b</mi> <mrow> <mo>(</mo> <msub> <mi>x</mi> <mi>i</mi> </msub> <mo>)</mo> </mrow> <mo>-</mo> <mi>u</mi> <mo>)</mo> </mrow> </mrow> </math>
in the formula, U is the number of color intervals, N is the number of pixel points in the target area, and a normalization constant Ch
<math> <mrow> <msub> <mi>C</mi> <mi>h</mi> </msub> <mo>=</mo> <mfrac> <mn>1</mn> <mrow> <munderover> <mi>&Sigma;</mi> <mrow> <mi>i</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>N</mi> </munderover> <mi>k</mi> <mrow> <mo>(</mo> <msub> <mi>x</mi> <mi>i</mi> </msub> <mo>)</mo> </mrow> </mrow> </mfrac> </mrow> </math>
2) Extracting edge characteristic information;
the method adopts an edge direction characteristic based on multi-scale block edge intensity weighting, and the steps are as follows:
(1) interpolating the target area to obtain an interpolation area 2 times as wide and as high as the target area, and then respectively blocking the target area and the interpolation area;
the method for interpolating the target region is known to those skilled in the art, and the blocking method uses overlapped blocks or non-overlapped blocks to divide the target region and the interpolation region into a plurality of sub-blocks, which is not limited in the embodiment of the present invention.
(2) Calculating the edge strength and the direction of each sub-block, dividing the edge direction into a plurality of direction areas within the range of 0-360 degrees, voting the direction areas according to the edge strength to obtain an edge feature histogram of each sub-block;
the edge strength and the edge direction are calculated by using the Sobel operator [2], other methods can be adopted, and the embodiment of the invention is not limited to this. The number of directional regions is set according to the needs of practical application, for example: the edge direction is divided into 18 direction areas in the range of 0-360 degrees with 20 degrees as an interval. The steps of voting for the direction area according to the edge strength are well known to those skilled in the art, and the embodiments of the present invention are not described herein.
(3) And connecting the edge feature histograms calculated by each sub-block to obtain a complete edge feature histogram.
And the dimension of the complete edge feature histogram is the product of the number of the direction areas and the total number of the sub-blocks.
3) Extracting texture feature information;
the method adopts a multi-scale local binary pattern [3] operator based on the blocks to calculate the texture characteristics, compared with the original local binary pattern characteristics, the multi-scale local binary pattern characteristics based on the blocks are less sensitive to the influence of image noise, richer local and global information can be extracted, and the method has stronger expression capability and discrimination capability on a target image and stronger robustness.
(1) Calculating a local binary pattern characteristic histogram for each sub-block;
wherein, each subblock is obtained by dividing the edge characteristic information, and the method adopts a unified mode [4] containing 59 patterns to calculate the local binary mode characteristics.
(2) And connecting the local binary pattern feature histograms calculated by each sub-block to obtain a complete texture feature histogram.
Wherein the dimension of the complete texture feature histogram is 59 times the total number of sub-blocks.
4) And fusing the color feature information, the edge feature information and the texture feature information to obtain a target feature histogram as a template feature.
Namely, color feature information, edge feature information and texture feature information are normalized, and the normalized three kinds of feature information are connected according to a preset weight (the weight value is set according to the requirement in practical application, for example, the weight ratio of the color feature information, the edge feature information and the texture feature information is 1: 1: 1), so as to obtain a target feature histogram.
102: initializing a recognizer;
the recognizer is used for finding the lost tracked target, the recognizer adopts Boost algorithm [5] training for learning, a positive sample in learning is a target area, a negative sample randomly selects a rectangular area with the same size as the positive sample in a background, the number of the negative samples is determined according to actual requirements, and the experimental reference value is 3.
103: inputting a next frame of image and initializing a candidate target area;
and according to Gaussian distribution, randomly selecting N areas around the target area as candidate target areas. The value of N is selected according to actual needs, and the reference value in the experiment is 20.
104: acquiring a new candidate target area according to a transfer formula;
Rn=A*(Rc-R0)+B*rng+R0
wherein R isn,Rc,R0Respectively a new candidate target area, a candidate target area and a target area; A. b is a transfer coefficient; rng is a random number generator. For example: and when the width parameter of the new candidate target area is calculated, the width parameters of the candidate target area and the width parameters of the target area are substituted into a transfer formula to obtain the width parameter of the new candidate target area, and by analogy, the height parameter and the center coordinate of the new candidate target area are respectively calculated, and finally the new candidate target area is obtained.
105: respectively extracting three visual characteristics of color, edge and texture from the new candidate target area;
the specific operation of this step is the same as that of step 101, and details thereof are not described in this embodiment of the present invention.
106: according to the discriminativity and the relativity of the color, the edge and the texture characteristics, carrying out self-adaptive fusion on the characteristics;
discriminative RDfThe degree of similarity between the new candidate target region and the adjacent background with respect to a certain feature is defined and is represented by the Papanicolaou coefficients of two histograms:
<math> <mrow> <msub> <mi>RD</mi> <mi>f</mi> </msub> <mo>=</mo> <mfrac> <mn>1</mn> <mrow> <munderover> <mi>&Sigma;</mi> <mrow> <mi>m</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>M</mi> </munderover> <msqrt> <msubsup> <mi>q</mi> <mrow> <mi>f</mi> <mrow> <mo>(</mo> <mi>bg</mi> <mo>)</mo> </mrow> </mrow> <mi>m</mi> </msubsup> <mo>&CenterDot;</mo> <msubsup> <mi>p</mi> <mi>f</mi> <mi>m</mi> </msubsup> </msqrt> </mrow> </mfrac> <mo>,</mo> <mi>f</mi> <mo>&Element;</mo> <mo>{</mo> <mi>HSV</mi> <mo>,</mo> <mi>EO</mi> <mo>,</mo> <mi>LBP</mi> <mo>}</mo> </mrow> </math>
in the formula, HSV, EO and LBP respectively represent color, edge and texture characteristics,andrespectively representing the m-th dimensions of the feature histograms of the adjacent background and the target area; m is the dimension of the color feature histogram or the dimension of the full edge feature histogram or the dimension of the full texture feature histogram. Discriminative RDfThe smaller the feature is, the stronger the distinguishing degree of the feature is, and the higher weight is assigned; instead, a lower weight is assigned.
The correlation is defined as the similarity degree of the new candidate target area and the template characteristic about a certain characteristic, and is represented by the Papanicolaou coefficients of two histograms:
<math> <mrow> <msub> <mi>CD</mi> <mi>f</mi> </msub> <mo>=</mo> <mfrac> <mn>1</mn> <mrow> <munderover> <mi>&Sigma;</mi> <mrow> <mi>m</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>M</mi> </munderover> <msqrt> <msubsup> <mi>q</mi> <mi>f</mi> <mi>m</mi> </msubsup> <mo>&CenterDot;</mo> <msubsup> <mi>p</mi> <mi>f</mi> <mi>m</mi> </msubsup> </msqrt> </mrow> </mfrac> <mo>,</mo> <mi>f</mi> <mo>&Element;</mo> <mo>{</mo> <mi>HSV</mi> <mo>,</mo> <mi>EO</mi> <mo>,</mo> <mi>LBP</mi> <mo>}</mo> </mrow> </math>
in the formula, HSV, EO and LBP respectively represent color, edge and texture characteristics,andrespectively representing the m-th dimension of the new candidate target area and the target area feature histogram; m is the dimension of the color feature histogram or the dimension of the complete edge feature histogram or the dimension of the complete texture feature histogram; correlation CDfThe larger the feature is, the stronger the correlation degree of the feature is, and the higher weight is assigned; instead, a lower weight is assigned.
The weights α, β, γ of each feature are ultimately determined by RDf and CDf:
<math> <mfenced open='{' close=''> <mtable> <mtr> <mtd> <mi>&alpha;</mi> <mo>=</mo> <mi>C</mi> <mo>&CenterDot;</mo> <msub> <mi>RD</mi> <mi>HSV</mi> </msub> <mo>&CenterDot;</mo> <msub> <mi>CD</mi> <mi>HSV</mi> </msub> </mtd> </mtr> <mtr> <mtd> <mi>&beta;</mi> <mo>=</mo> <mi>C</mi> <mo>&CenterDot;</mo> <msub> <mi>RD</mi> <mi>EO</mi> </msub> <mo>&CenterDot;</mo> <msub> <mi>CD</mi> <mi>EO</mi> </msub> </mtd> </mtr> <mtr> <mtd> <mi>&gamma;</mi> <mo>=</mo> <mi>C</mi> <mo>&CenterDot;</mo> <msub> <mi>RD</mi> <mi>LBP</mi> </msub> <mo>&CenterDot;</mo> <msub> <mi>CD</mi> <mi>LBP</mi> </msub> </mtd> </mtr> </mtable> </mfenced> </math>
in the formula,is a normalization constant, α + β + γ = 1; HSV, EO, LBP represent color, edge, texture features, respectively.
107: calculating the Pasteur distance d between the fused features and the template features, carrying out normalization processing on the Pasteur distance d, and taking the normalization result as the weight of a new candidate target region;
<math> <mrow> <mi>d</mi> <mo>=</mo> <msqrt> <mn>1</mn> <mo>-</mo> <mi>&rho;</mi> </msqrt> </mrow> </math>
<math> <mrow> <mi>&rho;</mi> <mo>=</mo> <munderover> <mi>&Sigma;</mi> <mrow> <mi>m</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>M</mi> </munderover> <msqrt> <msup> <mi>p</mi> <mi>m</mi> </msup> <mo>&CenterDot;</mo> <msup> <mi>q</mi> <mi>m</mi> </msup> </msqrt> </mrow> </math>
wherein ρ represents the babbitt coefficient; m is the dimension or the complete edge of the color feature histogramThe dimension of the feature histogram or the dimension of the complete texture feature histogram; q. q.sm,pmRespectively, the mth dimension of the fused feature and the template feature.
108: sorting the N new candidate target areas according to the weight, and calculating a resampling judgment value NjIf the resampling judgment value is larger than the resampling judgment threshold value, resampling is carried out, and step 109 is executed; if not, go to step 109;
wherein,n is the number of new candidate target areas;is the ith new candidate target region weight.
The resampling decision threshold is empirically obtained with a reference value of 50. Resampling, namely deleting the new candidate target area with small weight, and then copying the new candidate target area with the maximum weight to replace the deleted new candidate target area.
109: performing overlap judgment on the new candidate target region with the maximum weight and the target region, and if the overlap ratio is smaller than the overlap ratio threshold, performing step 110; otherwise, go to step 111;
the overlap ratio threshold is set according to the situation in practical application, for example: 0.3, which is not limited by the embodiments of the present invention when specifically implemented.
110: inputting multiple regions (width LW, height LH, center (x, y), L is multiple) of the new candidate target region (width W, height H, center (x, y)) with the largest weight into the detector, and if the detector outputs 0, indicating that the tracking fails; otherwise, inputting the output result of the detector into the recognizer, and if the output of the recognizer is yes, indicating that the tracking is successful and updating the recognizer, the template features and the target area; if the output is not, indicating that a new target is found, and ending the process;
the detector is a general classifier that can detect objects of interest to distinguish the tracked object from other objects. The detector is trained off-line and will not be updated during the tracking process. The detector input may be a frame image or a region in a frame image, and the output is the region of the target object in the image.
The identifier is used to obtain a priori information of a specific target in the tracking process, and is updated only in specific situations. The positive samples in the update of the identifier are composed of new candidate target areas verified by the detector, and the negative samples are randomly selected areas with the same size as the new candidate target areas in the background. The identifier prevents the occurrence of tracking drift by identifying a verified new candidate target region. The recognizer input is a target area and the output is a logical value indicating whether the target belongs to the target represented by the recognizer. Taking the feature of the new candidate target region with the maximum weight as the updated template feature; and taking the new candidate target area with the maximum weight as the updated target area.
111: and if the overlapping rate is more than or equal to the overlapping rate threshold value, the tracking is considered to be successful, the identifier, the template features and the target area are updated, and the process is ended.
The operation of updating the identifier, the template feature and the target area in this step is consistent with the updating operation in step 110, and the embodiment of the present invention is not described herein again.
Steps 104 through 111 are repeated for other frames of the video sequence until the entire video is traversed.
The feasibility of the target tracking method based on multi-feature adaptive fusion and online learning provided by the embodiment of the invention is verified by specific experiments, and the experiments track people as targets, and specific results are shown in fig. 2 to 7. The detector is a human body detector trained on edge gradient histogram features and a support vector machine model.
Fig. 2, fig. 3 and fig. 4 show the 16 th, 19 th and 23 th frames of a video sequence, respectively. In the 16 th frame, the target area is initialized, in the 19 th frame, the target 1 is shielded and lost, and in the 23 th frame, the shielded and lost target is successfully found out through online learning, so that the method is proved to be capable of effectively avoiding the problem of target loss after shielding.
Fig. 5, 6 and 7 are 291, 294 and 299 frames of a video sequence, respectively. In the 291 th frame, the target area is initialized, the 294 th frame, the target 1 and the target 2 are in staggered shielding, and the 299 th frame, the method carries out accurate and effective tracking, and avoids the problem that the tracking drift is easy to occur after the target is in staggered shielding.
Reference to the literature
[1]N.Dalal and B.Triggs.Histograms of oriented gradients for human detection.CVPR,2005.
[2]Kanopoulos N,Vasanthavada N,Baker R.L.Design of an image edge detectionfilter using the Sobel operator.IEEE Journal of Volume:23,Issue.1988.
[3]Wang Xiaoyu,Han,Tony X,Yan Shuicheng.An HOG-LBP human detector withpartial occlusion handling.Computer Vision,32-39,2009.
[4]T.Ahonen,A.Hadid,and M.Pietikainen.Face Recogniton with Local BinaryPatterns.Proc.European Conf.Computer Vision,469-481,2004.
[5]P.Viola and M.Jones.Rapid object detection using a boosted cascade of simplefeatures.In Proc.CVPR,volume I,511-518,2001.
Those skilled in the art will appreciate that the drawings are only schematic illustrations of preferred embodiments, and the above-described embodiments of the present invention are merely provided for description and do not represent the merits of the embodiments.
The above description is only for the purpose of illustrating the preferred embodiments of the present invention and is not to be construed as limiting the invention, and any modifications, equivalents, improvements and the like that fall within the spirit and principle of the present invention are intended to be included therein.

Claims (5)

1. A target tracking method based on multi-feature adaptive fusion and online learning is characterized by comprising the following steps:
(1) selecting a target area from a frame of image, extracting target characteristics and using the target characteristics as template characteristics;
(2) the initialization recognizer inputs a next frame of image and initializes a candidate target area; acquiring a new candidate target area according to a transfer formula;
(3) respectively extracting three characteristics of color, edge and texture from the new candidate target area; performing self-adaptive fusion according to the discriminativity and the correlation of the features;
(4) calculating the Bhattacharyya distance between the fused features and the template features, and normalizing the Bhattacharyya distance to be used as the weight of the new candidate target area;
(5) sorting the N new candidate target areas according to the weight, if the resampling judgment value is larger than the resampling judgment threshold value, resampling, and executing the step (6); if not, executing the step (6);
(6) performing overlapping judgment on the new candidate target area with the maximum weight and the target area, and executing the step (7) if the overlapping rate is smaller than an overlapping rate threshold value; otherwise, executing step (8);
(7) inputting the multiple regions of the new candidate target region with the largest weight into a detector, and if the detector outputs 0, indicating that the tracking fails; otherwise, inputting the output result of the detector into a recognizer, and if the output of the recognizer is yes, indicating that the tracking is successful and updating the recognizer, the template features and the target area; if the output is not, indicating that a new target is found, and ending the process;
(8) if the overlapping rate is greater than or equal to the overlapping rate threshold value, the tracking is considered to be successful, the identifier, the template features and the target area are updated, and the process is ended;
wherein,
the distinguishability is defined as the similarity degree of the new candidate target area and the adjacent background about a certain characteristic and is represented by the Papanicolaou coefficients of two histograms;
the correlation is defined as the similarity degree of the new candidate target area and the template characteristic about a certain characteristic and is represented by the Papanicolaou coefficient of two histograms;
the update identifier is specifically: the positive sample consists of a new candidate target area verified by the detector, and the negative sample is an area with the same size as the new candidate target area randomly selected from the background;
the updating template is characterized in that: taking the feature of the new candidate target region with the maximum weight as the updated template feature;
the update target area specifically includes: and taking the new candidate target area with the maximum weight as the updated target area.
2. The target tracking method based on the multi-feature adaptive fusion and the online learning according to claim 1, wherein the step of extracting the target feature and using the extracted target feature as the template feature specifically comprises:
1) extracting color characteristic information;
2) extracting edge characteristic information;
3) extracting texture feature information;
4) and fusing the color feature information, the edge feature information and the texture feature information to obtain a target feature histogram as the template feature.
3. The target tracking method based on the multi-feature adaptive fusion and the online learning according to claim 2, wherein the step of extracting the color feature information specifically comprises:
1) dividing a color space into a colored area and an achromatic area, carrying out HSV (hue, saturation and value) partition on the colored area and the achromatic area to obtain QH×QSA color subinterval and QvAn achromatic subinterval, dividing said QH×QSA color subinterval and said QvOne achromatic subinterval as QH×QS+QvEach color interval u;
2) endowing different weights to the pixel points according to the distance between the pixel points and the central point of the target area, and voting the corresponding color interval u according to the HSV of the pixel points;
3) and counting the voting value of each color interval to obtain a color feature histogram.
4. The target tracking method based on the multi-feature adaptive fusion and the online learning according to claim 2, wherein the step of extracting the edge feature information specifically comprises:
1) interpolating the target area to obtain an interpolation area 2 times as wide and as high as the target area, and then respectively blocking the target area and the interpolation area;
2) calculating the edge strength and the direction of each sub-block, dividing the edge direction into a plurality of direction areas within the range of 0-360 degrees, voting the direction areas according to the edge strength to obtain an edge feature histogram of each sub-block;
3) and connecting the edge feature histograms calculated by each sub-block to obtain a complete edge feature histogram.
5. The target tracking method based on the multi-feature adaptive fusion and the online learning according to claim 2, wherein the step of extracting the texture feature information specifically comprises:
1) calculating a local binary pattern characteristic histogram for each sub-block;
2) and connecting the local binary pattern feature histograms calculated by each sub-block to obtain a complete texture feature histogram.
CN201310121576.9A 2013-04-09 2013-04-09 A kind ofly to merge and the method for tracking target of on-line study based on multiple features self-adaptation Expired - Fee Related CN103198493B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201310121576.9A CN103198493B (en) 2013-04-09 2013-04-09 A kind ofly to merge and the method for tracking target of on-line study based on multiple features self-adaptation

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201310121576.9A CN103198493B (en) 2013-04-09 2013-04-09 A kind ofly to merge and the method for tracking target of on-line study based on multiple features self-adaptation

Publications (2)

Publication Number Publication Date
CN103198493A CN103198493A (en) 2013-07-10
CN103198493B true CN103198493B (en) 2015-10-28

Family

ID=48720998

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201310121576.9A Expired - Fee Related CN103198493B (en) 2013-04-09 2013-04-09 A kind ofly to merge and the method for tracking target of on-line study based on multiple features self-adaptation

Country Status (1)

Country Link
CN (1) CN103198493B (en)

Families Citing this family (17)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103473288B (en) * 2013-08-29 2016-12-28 西北工业大学 A kind of image search method describing son based on mixing micro structure
CN103996046B (en) * 2014-06-11 2017-07-21 北京邮电大学 The personal identification method merged based on many visual signatures
CN106033613B (en) * 2015-03-16 2019-04-30 北京大学 Method for tracking target and device
CN105069488B (en) * 2015-09-25 2018-06-29 南京信息工程大学 Tracking based on template on-line talking
CN105513090B (en) * 2015-11-24 2018-06-05 浙江宇视科技有限公司 A kind of observing and nursing construction method and device
CN105844647A (en) * 2016-04-06 2016-08-10 哈尔滨伟方智能科技开发有限责任公司 Kernel-related target tracking method based on color attributes
CN105957107A (en) * 2016-04-27 2016-09-21 北京博瑞空间科技发展有限公司 Pedestrian detecting and tracking method and device
CN106991395B (en) * 2017-03-31 2020-05-26 联想(北京)有限公司 Information processing method and device and electronic equipment
CN108876841B (en) * 2017-07-25 2023-04-28 成都通甲优博科技有限责任公司 Interpolation method and system in parallax refinement of parallax map
CN107633226B (en) * 2017-09-19 2021-12-24 北京师范大学珠海分校 Human body motion tracking feature processing method
CN108846854B (en) * 2018-05-07 2021-03-16 中国科学院声学研究所 Vehicle tracking method based on motion prediction and multi-feature fusion
CN110147768B (en) * 2019-05-22 2021-05-28 云南大学 Target tracking method and device
CN110232703B (en) * 2019-06-12 2023-07-25 中国矿业大学 Moving object recognition device and method based on color and texture information
CN110414535B (en) * 2019-07-02 2023-04-28 绵阳慧视光电技术有限责任公司 Manual initial frame correction method and system based on background distinction
CN110443320A (en) * 2019-08-13 2019-11-12 北京明略软件系统有限公司 The determination method and device of event similarity
CN110569840B (en) * 2019-08-13 2023-05-16 浙江大华技术股份有限公司 Target detection method and related device
CN114692788B (en) * 2022-06-01 2022-08-19 天津大学 Early warning method and device for extreme weather of Ernino based on incremental learning

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP2037410A1 (en) * 2007-09-14 2009-03-18 Thomson Licensing Method for tracking an object in a sequence of images and device implementing said method
CN102982340A (en) * 2012-10-31 2013-03-20 中国科学院长春光学精密机械与物理研究所 Target tracking method based on semi-supervised learning and random fern classifier
CN102999920A (en) * 2012-10-25 2013-03-27 西安电子科技大学 Target tracking method based on nearest neighbor classifier and mean shift

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP2037410A1 (en) * 2007-09-14 2009-03-18 Thomson Licensing Method for tracking an object in a sequence of images and device implementing said method
CN102999920A (en) * 2012-10-25 2013-03-27 西安电子科技大学 Target tracking method based on nearest neighbor classifier and mean shift
CN102982340A (en) * 2012-10-31 2013-03-20 中国科学院长春光学精密机械与物理研究所 Target tracking method based on semi-supervised learning and random fern classifier

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
一种基于多特征自适应融合的运动目标跟踪算法;尹宏鹏 等;《光电子·激光》;20100630;第21卷(第6期);3、4部分 *
基于多特征融合的粒子滤波目标跟踪算法;刘士荣 等;《信息与控制》;20120630;第41卷(第6期);752-759页 *

Also Published As

Publication number Publication date
CN103198493A (en) 2013-07-10

Similar Documents

Publication Publication Date Title
CN103198493B (en) A kind ofly to merge and the method for tracking target of on-line study based on multiple features self-adaptation
Hashemzadeh et al. Fire detection for video surveillance applications using ICA K-medoids-based color model and efficient spatio-temporal visual features
Mandal et al. AVDNet: A small-sized vehicle detection network for aerial visual data
Tsai et al. Road sign detection using eigen colour
Vignesh et al. Abnormal event detection on BMTT-PETS 2017 surveillance challenge
Masita et al. Pedestrian detection using R-CNN object detector
Halidou et al. Fast pedestrian detection based on region of interest and multi-block local binary pattern descriptors
Nandhini et al. SIFT algorithm-based Object detection and tracking in the video image
Naik Bukht et al. A Novel Human Interaction Framework Using Quadratic Discriminant Analysis with HMM.
Deshmukh et al. Real-time traffic sign recognition system based on colour image segmentation
Anjomshoae et al. Enhancement of template-based method for overlapping rubber tree leaf identification
Satsangi et al. License plate recognition: a comparative study on thresholding, OCR and machine learning approaches
Fritz et al. Object recognition using local information content
Kaur et al. 2-D geometric shape recognition using canny edge detection technique
Vijayalakshmi et al. Performance evaluation of object detection techniques for object detection
Eldho et al. YOLO based Logo detection
Oluchi et al. Development of a Nigeria vehicle license plate detection system
Phung et al. A new image feature for fast detection of people in images
Chen et al. Context-aware lane marking detection on urban roads
Chen et al. Human detection in surveillance video
Cui et al. Pedestrian detection using improved histogram of oriented gradients
Rao et al. Object Sub-Categorization and Common Framework Method using Iterative AdaBoost for Rapid Detection of Multiple Objects
Husain et al. Estimating the Number of People in Digital Still Images Based on Viola-Jones Face Detection Algorithms
Kerdvibulvech Hybrid model of human hand motion for cybernetics application
Baquhaizel et al. Person re-identification based on sad and histogram in a non-overlapping camera network

Legal Events

Date Code Title Description
C06 Publication
PB01 Publication
C10 Entry into substantive examination
SE01 Entry into force of request for substantive examination
C14 Grant of patent or utility model
GR01 Patent grant
CF01 Termination of patent right due to non-payment of annual fee
CF01 Termination of patent right due to non-payment of annual fee

Granted publication date: 20151028

Termination date: 20200409