CN110047090B - RGB-D target tracking method based on evolution feature learning - Google Patents
RGB-D target tracking method based on evolution feature learning Download PDFInfo
- Publication number
- CN110047090B CN110047090B CN201910240994.7A CN201910240994A CN110047090B CN 110047090 B CN110047090 B CN 110047090B CN 201910240994 A CN201910240994 A CN 201910240994A CN 110047090 B CN110047090 B CN 110047090B
- Authority
- CN
- China
- Prior art keywords
- rgb
- chromosome
- chromosomes
- depth
- image
- 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.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
- G06T7/0012—Biomedical image inspection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/20—Analysis of motion
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/50—Depth or shape recovery
- G06T7/55—Depth or shape recovery from multiple images
- G06T7/593—Depth or shape recovery from multiple images from stereo images
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10016—Video; Image sequence
- G06T2207/10021—Stereoscopic video; Stereoscopic image sequence
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10056—Microscopic image
Landscapes
- Engineering & Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Physics & Mathematics (AREA)
- Medical Informatics (AREA)
- Quality & Reliability (AREA)
- Radiology & Medical Imaging (AREA)
- Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses an RGB-D target tracking method based on evolution feature learning, which applies an evolution feature learning technology to an RGB-D video target tracking method, does not need prior knowledge of a target in the evolution feature learning, and can overcome the defect of manually designing the feature. Meanwhile, the number of parameters for the evolutionary feature learning is much smaller than that of the deep neural network, and a large amount of time is not needed for learning and training the parameters. Therefore, the method can overcome the technical defects of poor robustness of the manually designed features and the technical defects of low calculation efficiency and long time consumption of a deep learning method.
Description
Technical Field
The invention relates to the field of artificial intelligence and machine vision, in particular to an RGB-D target tracking method based on evolution feature learning.
Background
In recent years, RGB-D video target tracking is a hot research problem in the field of artificial intelligence, and has very wide application in the fields of virtual reality, human-computer interaction, unmanned driving and the like.
However, most of the existing target tracking methods only use RGB data to extract features, which causes a certain limitation on tracking performance, and when the illumination condition changes, the target is blocked, and the appearance of the target changes, the tracking accuracy decreases. In the other part of the method, RGB-D data is adopted for target tracking, and the method is roughly divided into two types, wherein one type is based on artificial design characteristics, and the other type is based on Deep learning characteristics; in which the manual design of features requires a lot of a priori knowledge, and the person who needs feature modeling has a very intensive study in this field, even though it is impossible to solve various possible difficulties in one model. The Deep learning characteristics can be used for learning and training a large amount of target characteristics automatically, so that the defects of artificial design characteristics are overcome to a great extent, but a large amount of parameters in a Depth neural network need to be learned and trained, which causes lower calculation efficiency and longer time consumption.
Disclosure of Invention
The purpose of the invention is as follows: the invention aims to provide an RGB-D target tracking method based on evolution feature learning, which has the advantages of no need of prior knowledge of targets and overcomes the defects of artificial design features and the technical defects of low calculation efficiency and long time consumption of a Deep learning method.
The technical scheme is as follows: the invention relates to an RGB-D target tracking method based on evolution feature learning, which comprises the following steps:
step 2, encoding chromosomes of candidate target images in RGB, and encoding chromosomes of candidate target images in Depth images; wherein one chromosome comprises 32 genes, each gene comprising 7 data items, one data item comprising 8 binary bits; in the RGB image, 7 data items respectively represent the center point coordinates (X, Y) of the candidate target image block, the width and height (W, H) of the image block, and the numerical values of the three channels R, G, B; in a Depth image, 7 data items respectively represent the central point coordinates (X, Y) of a candidate target image block, the width and height (W, H) of the image block, the horizontal difference, the height to the ground and the numerical values of three channels of the angle of a surface normal vector;
step 3, carrying out initialization candidate solution space on chromosomes of the candidate target image in the RGB and Depth image modes, wherein the population of the chromosomes in the initialization candidate solution space is represented as:or Is a chromosome, j =1,2, …, M is the number of chromosomes contained in the population, G =1,2, …, G represents the number of generations;
calculating the matching degree of each candidate target image and the last frame tracking result in the candidate solution space, wherein the matching degree is calculated as X for each input image in RGB or Depth mode RGB Or X depth In the evolution algorithm, there is one reconstructed imageOrC j Representing the jth chromosome, for the candidate chromosomes, the corresponding degree of match equation is:
and taking the matched chromosome as a termination result, and simultaneously outputting the evolution characteristics of the chromosome in the RGB and depth modes.
Further, there is a chromosome with the highest matching degree among the chromosomes in the RGB and depth patterns, and the chromosome with the highest matching degree can be obtained by the following formula:
derived fromAs X RGB Of the evolving characteristics of (2) and final results F RGB ,As X depth Of the evolving characteristics of (2) and final results F depth 。
Further, in step 3, chromosomes which do not satisfy the termination result are subjected to further selection, including combinatorial crossing, mutation, generation of new populations, and the above process is repeated.
Further, in the selection process, each chromosome is ranked according to the matching degree of the chromosome, the worst is rank1, next rank 2 … … is preferably rank M, and M is the number of chromosomes contained in the population; calculating the selection probability according to the rank value of the chromosome:
P j to select the probability of the j-th chromosome,is the probability of selecting the worst chromosome,is the probability of selecting the best chromosome, m worst M is the number of chromosomes with the lowest degree of matching best M is the number of chromosomes with the highest degree of matching worst M is the number of chromosomes with the lowest degree of matching best The number of chromosomes with the highest matching degree.
Further, the matching chromosomes in the RGB and Depth modes can be used to analyze and obtain the error amount of the evolution features between two adjacent frames, which is according to the following formula:
wherein, F RGB Is X RGB End result of the evolving characteristics of (1), F depth Is X depth I, i is the number of sequences of the frame.
Has the advantages that: the method applies the evolution characteristic learning technology to the RGB-D video target tracking method, the evolution characteristic learning does not need the prior knowledge of the target, and the defect of manually designing the characteristic can be overcome. Meanwhile, the number of parameters for evolving feature learning is much smaller than that of a deep neural network, and a large amount of time is not needed for learning and training the parameters. Therefore, the method can overcome the technical defects of poor robustness of the manually designed features and the technical defects of low calculation efficiency and long time consumption of a deep learning method.
Drawings
FIG. 1 is a schematic representation of chromosome coding in RGB and Depth modes;
FIG. 2 is a flow chart of calculating evolution characteristics in RGB and Depth modes;
FIG. 3 is a flow chart of an RGB-D tracking method based on evolutionary feature learning;
fig. 4 is a diagram illustrating a part of tracking effects in the present invention.
Detailed Description
As shown in fig. 1 and 2, an RGB-D target tracking method based on evolutionary feature learning includes the following steps:
step 2, encoding chromosomes of candidate target images in RGB, and encoding chromosomes of candidate target images in Depth images; wherein one chromosome comprises 32 genes, each gene comprising 7 data items, one data item comprising 8 binary bits; in the RGB image, 7 data items respectively represent the center point coordinates (X, Y) of the candidate target image block, the width and height (W, H) of the image block, and the numerical values of the three channels R, G, B; in a Depth image, 7 data items respectively represent the central point coordinates (X, Y) of a candidate target image block, the width and height (W, H) of the image block, the horizontal difference, the height to the ground and the numerical values of three channels of the angle of a surface normal vector;
step 3, chromosome of candidate target image under RGB and Depth image mode is initialized to candidate solution space, and initialThe population of chromosomes in the solution space candidate is represented as:or Is a chromosome, j =1,2, …, M is the number of chromosomes contained in the population, G =1,2, …, G represents the number of generations;
calculating the matching degree of each candidate target image and the last frame tracking result in the candidate solution space, wherein the matching degree is calculated as X for each input image in an RGB or Depth mode RGB Or X depth In the evolution algorithm, there is one reconstructed imageOr alternativelyC j Representing the jth chromosome, for the candidate chromosomes, the corresponding degree of match equation is:
and taking the matched chromosome as a termination result, and simultaneously outputting the evolution characteristics of the chromosome in the RGB and depth modes.
Among chromosomes in the RGB and depth patterns, there is a chromosome with the highest matching degree, which can be obtained by the following formula:
derived fromAs X RGB Of the evolving characteristics of (2) and final results F RGB ,As X depth Of the evolving characteristics of (2) and final results F depth 。
In step 3, chromosomes which do not meet the termination result are further selected, including combination crossing and mutation, so as to generate a new population, and then the process is circulated.
In the selection process, each chromosome is sorted according to the matching degree of the chromosome, the worst is rank1, next rank 2 … … is most preferably rank M, and M is the number of chromosomes contained in the population; calculating the selection probability according to the rank value of the chromosome:
P j to select the probability of the j-th chromosome,is the probability of selecting the worst chromosome,is the probability of selecting the best chromosome, m worst M is the number of chromosomes with the lowest degree of matching best M is the number of chromosomes with the highest degree of matching worst M is the number of chromosomes with the lowest degree of matching best The number of chromosomes with the highest matching degree.
As shown in fig. 3, the matching chromosomes in RGB and Depth modes can be used to analyze and obtain the error amount of the evolution features between two adjacent frames, which is according to the following formula:
wherein, F RGB Is X RGB End result of the evolving characteristics of (1), F depth Is X depth T =1, 2.. I, i is the number of sequences of the frame.
The experimental results and analyses were as follows:
we used the Windows 10 operating system with MATLAB R2016a as the software platform. The main configuration of the computer is a 12-core processor, 54GB ram, nvidia GeForce GT650m GPU, the population contains the number of chromosomes M =30, the number of generations G =200, and the number of samples N =100, experimental performance was tested on a PTB data set, which contains 100 RGB-D videos, and the tracking success rate results are shown in table 1:
table 1.
Claims (5)
1. An RGB-D target tracking method based on evolution feature learning is characterized by comprising the following steps:
step 1, HHA coding is carried out on a Depth image of a candidate target in video target tracking;
step 2, encoding chromosomes of candidate target images in RGB (red, green and blue) and encoding chromosomes of candidate target images in a Depth image; wherein one chromosome comprises 32 genes, each gene comprising 7 data items, one data item comprising 8 binary bits; in the RGB image, 7 data items respectively represent the center point coordinates (X, Y) of the candidate target image block, the width and height (W, H) of the image block, and the numerical values of the three channels R, G, B; in a Depth image, 7 data items respectively represent central point coordinates (X, Y) of candidate target image blocks, width and height (W, H) of the image blocks, horizontal difference, ground height and the numerical values of three channels of angles of surface normal vectors;
step 3, carrying out initialization candidate solution space on chromosomes of the candidate target image in the RGB and Depth image modes, wherein the population of the chromosomes in the initialization candidate solution space is represented as:or Is a chromosome, j =1,2, …, M is the number of chromosomes contained in the population, G =1,2, …, G indicates the number of generations;
calculating the matching degree of each candidate target image and the last frame tracking result in the candidate solution space, wherein the matching degree is calculated as X for each input image in RGB or Depth mode RGB Or X depth In the evolution algorithm, there is one reconstructed imageOrC j Representing the jth chromosome, for the candidate chromosomes, the corresponding degree of match equation is:
And taking the matched chromosome as a termination result, and simultaneously outputting the evolution characteristics of the chromosome in the RGB and depth modes.
2. The RGB-D target tracking method based on evolutionary feature learning as claimed in claim 1, wherein in step 3, there exists a chromosome with the highest matching degree among the chromosomes in RGB and depth modes, and the chromosome with the highest matching degree can be obtained by the following formula:
3. The method for RGB-D object tracking based on evolutionary feature learning as claimed in claim 1, wherein in step 3, chromosomes that do not satisfy termination results are further selected, including combining crossover, mutation, generating new population, and then repeating the above process.
4. The RGB-D target tracking method based on evolutionary feature learning as claimed in claim 3, wherein in the selection process, each chromosome is ranked according to its matching degree, the worst is rank1, next rank 2 … … is the best is rank M, M is the number of chromosomes contained in the population; calculating the selection probability according to the rank value of the chromosome:
5. The RGB-D target tracking method based on evolving feature learning as claimed in claim 2, wherein the matching chromosomes in RGB and Depth modes can be used to analyze and obtain the error amount of evolving features between two adjacent frames, the error amount is according to the following formula:
wherein, F RGB Is X RGB End result of the evolving characteristics of (1), F depth Is X depth I, i is the number of sequences of the frame.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910240994.7A CN110047090B (en) | 2019-03-28 | 2019-03-28 | RGB-D target tracking method based on evolution feature learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910240994.7A CN110047090B (en) | 2019-03-28 | 2019-03-28 | RGB-D target tracking method based on evolution feature learning |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110047090A CN110047090A (en) | 2019-07-23 |
CN110047090B true CN110047090B (en) | 2022-10-14 |
Family
ID=67275387
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910240994.7A Active CN110047090B (en) | 2019-03-28 | 2019-03-28 | RGB-D target tracking method based on evolution feature learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110047090B (en) |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104809229A (en) * | 2015-05-07 | 2015-07-29 | 北京京东尚科信息技术有限公司 | Method and system for extracting text characteristic words |
CN107992827A (en) * | 2017-12-03 | 2018-05-04 | 湖南工程学院 | A kind of method and device of the multiple mobile object tracking based on threedimensional model |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US6957200B2 (en) * | 2001-04-06 | 2005-10-18 | Honeywell International, Inc. | Genotic algorithm optimization method and network |
-
2019
- 2019-03-28 CN CN201910240994.7A patent/CN110047090B/en active Active
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104809229A (en) * | 2015-05-07 | 2015-07-29 | 北京京东尚科信息技术有限公司 | Method and system for extracting text characteristic words |
CN107992827A (en) * | 2017-12-03 | 2018-05-04 | 湖南工程学院 | A kind of method and device of the multiple mobile object tracking based on threedimensional model |
Non-Patent Citations (2)
Title |
---|
Genetic Algorithm for Depth Images;Danciu G,et al.;《2014 IEEE 20th International Symposium for Design and Technology in Electronic Packaging》;20141026;全文 * |
Visual tracking with genetic algorithm augmented;Qu L,et a1.;《Signal Image&Video Processing》;20170601;全文 * |
Also Published As
Publication number | Publication date |
---|---|
CN110047090A (en) | 2019-07-23 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Man et al. | Genetic algorithms: concepts and designs | |
CN101271572B (en) | Image segmentation method based on immunity clone selection clustering | |
CN106529574B (en) | Image classification method based on sparse autocoder and support vector machines | |
CN113469356A (en) | Improved VGG16 network pig identity recognition method based on transfer learning | |
CN109446185B (en) | Collaborative filtering missing data processing method based on user clustering | |
CN107330902B (en) | Chaotic genetic BP neural network image segmentation method based on Arnold transformation | |
CN103942571B (en) | Graphic image sorting method based on genetic programming algorithm | |
CN111898689A (en) | Image classification method based on neural network architecture search | |
CN111832949A (en) | Construction method of equipment combat test identification index system | |
de Oliveira et al. | Image segmentation using multilevel thresholding and genetic algorithm: an approach | |
CN114219824A (en) | Visible light-infrared target tracking method and system based on deep network | |
CN107451617B (en) | Graph transduction semi-supervised classification method | |
CN114743027A (en) | Weak supervision learning-guided cooperative significance detection method | |
Goel et al. | Gray level enhancement to emphasize less dynamic region within image using genetic algorithm | |
CN114241267A (en) | Structural entropy sampling-based multi-target architecture search osteoporosis image identification method | |
CN110047090B (en) | RGB-D target tracking method based on evolution feature learning | |
CN113269617A (en) | Large-scale multi-objective optimization commodity recommendation method based on image coding | |
CN110047088B (en) | HT-29 image segmentation method based on improved teaching and learning optimization algorithm | |
CN116051924B (en) | Divide-and-conquer defense method for image countermeasure sample | |
CN111814884A (en) | Target detection network model upgrading method based on deformable convolution | |
CN116306919A (en) | Large-scale multi-objective combination optimization method based on problem recombination and application | |
CN114581470B (en) | Image edge detection method based on plant community behaviors | |
CN113469270B (en) | Semi-supervised intuitive clustering method based on decomposition multi-target differential evolution superpixel | |
CN106709572A (en) | Data processing method and equipment | |
CN108090514B (en) | Infrared image identification method based on two-stage density clustering |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant | ||
TR01 | Transfer of patent right |
Effective date of registration: 20221220 Address after: Room 305 and 307, No. 68 Shengtai Road, Jiangning Development Zone, Nanjing, Jiangsu 211106 Patentee after: Nanjing Longyuan Zhongchuang Information Technology Co.,Ltd. Address before: 223005 A12-2, high tech Industrial Park, three East seven street, Hongze District, Huaian, Jiangsu (Hongze technology transfer center Hongze sub center) Patentee before: HUAIYIN INSTITUTE OF TECHNOLOGY |
|
TR01 | Transfer of patent right |