CN109255339B - Classification method based on self-adaptive deep forest human gait energy map - Google Patents

Classification method based on self-adaptive deep forest human gait energy map Download PDF

Info

Publication number
CN109255339B
CN109255339B CN201811222012.3A CN201811222012A CN109255339B CN 109255339 B CN109255339 B CN 109255339B CN 201811222012 A CN201811222012 A CN 201811222012A CN 109255339 B CN109255339 B CN 109255339B
Authority
CN
China
Prior art keywords
forest
sample
adaptive
training
adaptive depth
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
Application number
CN201811222012.3A
Other languages
Chinese (zh)
Other versions
CN109255339A (en
Inventor
赵盼盼
盛立杰
苗启广
马悦
庞博
秦丹
陈红颖
徐劲夫
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.)
Xidian Univ
Original Assignee
Xidian Univ
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 Xidian Univ filed Critical Xidian Univ
Priority to CN201811222012.3A priority Critical patent/CN109255339B/en
Publication of CN109255339A publication Critical patent/CN109255339A/en
Application granted granted Critical
Publication of CN109255339B publication Critical patent/CN109255339B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING; COUNTING
    • G06KRECOGNITION OF DATA; PRESENTATION OF DATA; RECORD CARRIERS; HANDLING RECORD CARRIERS
    • G06K9/00Methods or arrangements for reading or recognising printed or written characters or for recognising patterns, e.g. fingerprints
    • G06K9/00335Recognising movements or behaviour, e.g. recognition of gestures, dynamic facial expressions; Lip-reading
    • G06K9/00342Recognition of whole body movements, e.g. for sport training
    • G06K9/00348Recognition of walking or running movements, e.g. gait recognition
    • GPHYSICS
    • G06COMPUTING; CALCULATING; COUNTING
    • G06NCOMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computer systems based on biological models
    • G06N3/004Artificial life, i.e. computers simulating life
    • G06N3/006Artificial life, i.e. computers simulating life based on simulated virtual individual or collective life forms, e.g. single "avatar", social simulations, virtual worlds or particle swarm optimisation

Abstract

The invention discloses a classification method based on a self-adaptive deep forest human gait energy map, which comprises the following steps: (1) constructing a training set and a test set; (2) constructing a self-adaptive depth forest model; (3) calculating the classification precision of the self-adaptive depth forest model; (4) judging whether the self-adaptive depth forest model meets the requirements or not; (5) if the accuracy and the depth of the training set do not meet the requirements, expanding the adaptive depth forest model, solving a user-defined secondary convex optimization problem to obtain the enhancement features of the adaptive depth forest model in the training set, and obtaining the training set of the expanded adaptive depth forest through the enhancement features on the training set; (6) and calculating the classification accuracy of the test set. The invention reduces the complexity of the model and the weight of the model needing to be trained, and simultaneously reduces the scale of the gait energy pattern needed by training the model.

Description

Classification method based on self-adaptive deep forest human gait energy map
Technical Field
The invention belongs to the technical field of image processing, and further relates to a human body gait energy map classification method based on self-adaptive depth forest in the technical field of image recognition. The invention can be used for classifying the gait characteristics in the human body gait Energy image GEI (Gate Energy image).
Background
The gait recognition technology is a biological technology for carrying out identity recognition according to the walking posture of a person in a video sequence. Because the gait recognition has the characteristics of non-invasiveness, long-distance recognition and difficulty in hiding, the gait recognition has wide application prospects in the fields of national public safety, financial safety, identity authentication, video monitoring and the like.
The patent document "gait recognition method based on deep learning" (patent application No. 201410587758X, application publication No. CN104299012A) applied by the limited corporation of the river water droplet science and technology (beijing) proposes a gait recognition method based on deep learning. The method adopts a gait energy diagram to describe a gait sequence, and trains a matching model through a deep convolutional neural network so as to identify the identity of a person by matching gait. The training process of the method comprises the following steps: extracting a gait energy diagram from a training gait video sequence with marked identities and related to a plurality of visual angles, and repeatedly selecting any two of the gait energy diagrams to train a matching model based on a convolutional neural network until the model converges; the identification process of the method is as follows: respectively extracting gait energy images of the single-view to-be-identified gait video sequence and the registered gait video sequence, calculating the similarity between the gait energy image of the single-view to-be-identified gait video sequence and each gait energy image of the registered gait video sequence by using a trained matching model based on a convolutional neural network in the training process, carrying out identity prediction according to the similarity, and outputting an identification result. The method has strong robustness to gait change across a large visual angle, and effectively solves the problem of low precision when the existing gait recognition technology is used for processing cross-visual angle gait recognition. However, the method still has the disadvantages that the method is a classification method based on the deep convolutional neural network, and the classification method based on the deep convolutional neural network needs a large number of training samples, and under the condition that the number of the training samples is small, the classification method based on the deep convolutional neural network cannot fully learn the gait features in the human gait energy diagram, so that the feature extraction capability of the deep convolutional neural network is reduced, the classification method based on the deep convolutional neural network is not suitable for data of small samples, the classification performance based on the deep convolutional neural network is seriously dependent on a parameter debugging process, and a large amount of computing resources are consumed in the training process.
A gait recognition method is provided in the patent document 'gait identity recognition method and system' (patent application No. CN201710136803.3, application publication No. CN107016346A) applied by the research of the Chinese academy of sciences computing technology. The method comprises the steps of extracting behavior characteristics, position characteristics and gait characteristics from acceleration data acquired by an intelligent terminal carried by a user; recognizing the current behavior of the user according to the behavior characteristics by utilizing a pre-trained behavior recognition model; identifying the current position of the intelligent terminal according to the position characteristics and the identified current behavior of the user by using a pre-trained position identification model; and identifying the identity of the user according to the gait characteristics, the identified current behavior of the user and the current position of the intelligent terminal by utilizing a pre-trained gait identification model. The method improves the accuracy and robustness of gait identification through a layered progressive identification mode, does not need to limit the placement position and direction of related sensors, and is flexible and convenient to use. However, the method still has the disadvantages that the method needs the user to carry an intelligent terminal to obtain acceleration data, and behavior characteristics, position characteristics and gait characteristics of the user are extracted by using the obtained acceleration data, so that the non-invasiveness of the gait classification method is weakened, and the method consists of three models, namely a pre-trained behavior recognition model, a pre-trained position recognition model and a pre-trained gait recognition model, so that the characteristic extraction cannot be completely integrated into one model, and the end-to-end classification cannot be realized.
Disclosure of Invention
The invention aims to provide a classification method based on a self-adaptive deep forest human gait energy map, aiming at the defects of the prior art. Compared with other existing gait classification methods, the adaptive deep forest model obtains the enhanced features of the adaptive deep forest model by solving the weight of each tree, and the adaptive deep forest model is trained by using the obtained enhanced features, so that the weights required by training are less, and the training speed is high. And the depth of the self-adaptive depth forest is continuously increased according to the accuracy of the test set and the training set in the training process, so that the complexity of the self-adaptive depth forest is adaptively adjusted along with the training process, the complexity of the model is reduced, and the computing resources are fully utilized.
The idea of realizing the purpose of the invention is that firstly, a gait energy graph is synthesized for all gait videos, and a training set and a test set are obtained by utilizing a random combination sampling method. And then constructing an initial adaptive depth forest model, and initializing the parameters of the model. And calculating the classification precision of the self-adaptive depth forest model of the previous iteration. Judging whether the classification precision of the current iteration adaptive depth forest model on a training set meets the requirement, stopping training if the classification precision meets the requirement, wherein the model is the required model, calculating the weight of each tree in the depth forest if the classification precision does not meet the requirement, obtaining the enhancement features of the current iteration adaptive depth forest model on a test set, stopping training if the accuracy on the test set meets the requirement, wherein the model is the required model, otherwise, expanding the current depth forest, obtaining the enhancement features of the training set of the depth forest, updating the training set of the current depth forest, increasing the depth of the depth forest, and continuing training the current depth forest until the required model is obtained.
The method comprises the following specific steps:
(1) constructing a sample set:
(1a) carrying out background modeling and background real-time updating on 124 input target character videos to obtain a human body outline foreground image of each target character;
(1b) extracting a human body gait image from the human body outline foreground image of each target person video;
(1c) adding the gray values of all the human body contour images of each target person obtained in each gait cycle, and dividing the sum by the total number of the human body contour images of each target person to obtain a gait energy map corresponding to the gait cycle;
(1d) taking six gait energy maps of each target person under a normal walking condition and with a visual angle of 90 degrees, and forming a sample set by 744 gait energy maps;
(2) obtaining a training set and a test set:
(2a) randomly sampling a sample set by using a random combination sampling splicing method to obtain 992 positive samples and negative samples, and forming a training set by using all the positive samples and the negative samples;
(2b) randomly sampling a sample set by using a random combination sampling splicing method to obtain 992 positive samples and negative samples, and forming a test set by using all the positive samples and the negative samples;
(3) constructing an adaptive depth forest model:
(3a) building a self-adaptive depth forest model consisting of two random forests and two completely random forests;
(3b) setting the number of trees in two random forests and two complete random forests in the self-adaptive depth forest model to be 400, setting the maximum depth of the trees to be 8, setting bootstrap sample marks to be True, setting the number of parallel-processed cores to be the maximum number of cores of a server, and setting the depth of the self-adaptive depth forest model to be h;
(4) calculating the classification precision of the current iteration self-adaptive depth forest model:
(4a) inputting a training set of the current iteration self-adaptive depth forest model into the current iteration self-adaptive depth forest model, and outputting class probability corresponding to each sample in the training set;
(4b) calculating the classification precision of the adaptive depth forest model after the training set is trained by the current iterative adaptive depth forest model by using a training set classification precision calculation formula;
(5) judging whether the classification precision of the current iteration self-adaptive depth forest model is greater than or equal to 95%; if yes, executing the step (12), otherwise, executing the step (6);
(6) obtaining the weight of each tree in the current iteration of the adaptive depth forest model:
(6a) obtaining the weight of each tree in each random forest of the current iterative self-adaptive depth forest model by using the following self-defined quadratic convex optimization expression:
ξi k≥0,i∈K
wherein the content of the first and second substances,expressing xi when minimizing an objective functioni,wkI denotes the index of the samples in the training set, K denotes the total number of samples in the training set, Σ denotes the summation operation, ξi kRepresenting the basic loss of the ith sample in the training set in the kth random forest in the current iterative adaptive depth forest model, lambda representing a regularization coefficient, | | | | - | representing 2 norm operation, wkRepresenting the total weight of the kth forest in the adaptive depth forest model of the current iteration, s.t. representing a constraint condition symbol, T representing a subscript of a tree in the kth random forest in the adaptive depth forest model of the current iteration, and TkRepresenting the total number of trees, P, in the kth random forest in the adaptive depth forest model for the current iterationi (t,k)Representing the fundamental loss, w, of the ith sample in the t tree of the kth random forest in the adaptive depth forest model of the current iteration(t,k)Representing the weight of the t number of the kth random forest in the current iteration of the adaptive depth forest model, and representing the classification interval by tau;
(6b) the weight of each tree in each random forest of the current iteration adaptive depth forest model is formed into the weight of the current iteration adaptive depth forest model;
(7) obtaining the enhancement characteristics of the current iteration self-adaptive depth forest model on the test set:
(7a) according to the weight of each tree in the current iteration self-adaptive depth forest model, respectively obtaining the enhancement features of each sample in the test set, namely a positive sample and a negative sample, through an enhancement feature solving formula of the current iteration self-adaptive depth forest model;
(7b) the enhancement features of the current iteration self-adaptive depth forest model are formed by the enhancement features of the positive samples and the negative samples of all samples in the test set;
(8) and (3) calculating the accuracy of the test set:
calculating the accuracy of the test set by using the accuracy calculation method of the adaptive depth forest model test set of the current iteration according to the enhancement characteristics of the adaptive depth forest model of the current iteration on the test set;
(9) judging whether the accuracy of the current iterative adaptive depth forest model test set is greater than the maximum accuracy of the test set, if so, executing the step (10), otherwise, executing the step (12);
(10) obtaining the enhancement characteristics of the current iteration self-adaptive depth forest model on a training set:
(10a) respectively obtaining the enhancement features of each sample in the training set, namely a positive sample and a negative sample, through an enhancement feature solving formula of the current iterative adaptive depth forest model according to the weight of each tree in the current iterative adaptive depth forest model;
(10b) the enhancement features of the current iteration self-adaptive depth forest model are formed by the enhancement features of the positive samples and the negative samples of all samples in the training set;
(11) expanding the self-adaptive depth forest model of the current iteration:
(11a) splicing the training set of the current iteration self-adaptive depth forest model and the enhancement features of the current iteration self-adaptive depth forest model end to obtain the training set of the current iteration self-adaptive depth forest;
(11b) taking the depth h +1 of the current iteration self-adaptive depth forest model as the depth h of the current iteration self-adaptive depth forest model;
(12) judging whether the depth h of the current iterative adaptive depth forest model is greater than 100, if so, executing the step (13) after obtaining the trained adaptive depth forest model, otherwise, executing the step (4);
(13) classifying all gait energy maps in the test set:
and inputting the test set of the current iterative adaptive depth forest model into the trained adaptive depth forest model, and classifying the gait energy map in the test set.
Compared with the prior art, the invention has the following advantages:
firstly, because the invention obtains the enhanced features of the adaptive deep forest model by solving the weight of each tree, and trains the adaptive deep forest model by utilizing the enhanced features, the invention overcomes the problems that a large amount of weights need to be trained in the process of training a neural network in the prior art, and the neural network model is not applicable to a small sample data set, so that the invention has the advantages of less trained weights, high training speed and suitability for a small data set.
Secondly, the depth of the self-adaptive deep forest is continuously increased according to the accuracy of the test set and the training set in the training process, so that the complexity of the self-adaptive deep forest is adaptively adjusted along with the training process, the problem that the complexity of a model is fixed and unchanged in the training process of a neural network in the prior art is solved, the complexity of the model is reduced, and computing resources are fully utilized.
Drawings
FIG. 1 is a flow chart of the present invention.
Detailed Description
The specific steps of the implementation of the present invention will be further described with reference to fig. 1.
Step 1, constructing a sample set.
And performing background modeling and background real-time updating on the input 124 target character videos to obtain a human body outline foreground image of each target character.
The background modeling and background real-time updating means that a previous frame and a background frame of a character target video are subtracted to obtain a human body outline foreground image of the target character video; and calling a function in a visual library to remove holes and scattered white points existing in the human body contour foreground image.
And extracting a human body gait image from the human body outline foreground image of each target person video.
The step of extracting the human body gait image is that the distance between the leftmost point and the rightmost point of each human body contour is taken as the width of the human body contour, and the distance between the uppermost point and the bottommost point is taken as the height of the human body contour; the gait cycle is divided by the width change signal of the human body contour by utilizing the characteristic that the contour width of the human body is synchronously and periodically changed along with time; and (3) scratching out the human body outline image according to the width and the height of the human body outline, keeping the height-width ratio unchanged, and scaling all the human body outline images to the same size.
And respectively adding the gray values of all the human body contour images of each target person obtained in each gait cycle, and then dividing the sum by the total number of the human body contour images of each target person to obtain a gait energy map corresponding to the gait cycle.
And taking six gait energy maps of each target person under the normal walking condition and with the visual angle of 90 degrees, and forming a sample set by 744 gait energy maps.
And 2, obtaining a training set and a testing set.
And randomly sampling the sample set by using a random combination sampling splicing method to obtain 992 positive samples and negative samples, and forming a training set by using all the positive samples and the negative samples.
And randomly sampling the sample set by using a random combination sampling splicing method to obtain 992 positive samples and negative samples, and forming a test set by using all the positive samples and the negative samples.
The random combination sampling splicing method comprises the following steps:
in the first step, 6 gait energy maps of an unselected human target are randomly selected from 124 human targets in the sample set as the current processing object.
And secondly, randomly selecting four pairs of gait energy maps from the current processing object, transversely splicing the gait energy maps of each pair to form a positive sample, setting the label of the positive sample to be 0, and setting the number of the positive samples of each target person to be 4.
And thirdly, randomly selecting a gait energy map from the current processing object.
And fourthly, respectively and randomly selecting 4 different target characters from the sample set except the current target character, and respectively and randomly selecting one gait energy image from 6 gait energy images of the selected target character, wherein 4 gait energy images are total.
And fifthly, pairing every two of the gait energy diagram obtained in the third step with the 4 state energy diagrams obtained in the fourth step to form 4 pairs of gait energy diagrams, transversely splicing the gait energy diagrams of each pair to form a negative sample, setting the label of the negative sample to be 1, and setting the number of the negative samples of each character target to be 4.
And a sixth step: and marking the current processing object as processed.
The seventh step: if 6 gait energy maps of each target person are marked as processed for 124 target persons, 496 positive samples and 496 negative samples are obtained in total, namely 992 samples in total, and if 6 gait energy maps of the target persons exist for 124 target persons, the first step is executed.
And 3, constructing a self-adaptive depth forest model.
And building a self-adaptive depth forest model consisting of two random forests and two completely random forests.
The number of trees in two random forests and two complete random forests in the self-adaptive depth forest model is set to be 400, the maximum depth of the trees is set to be 8, bootstrap sample flags are set to be True, the number of parallel processed cores is set to be the maximum number of cores of the server, and the depth of the self-adaptive depth forest model is set to be t.
And 4, calculating the classification precision of the current iterative adaptive depth forest model.
And inputting the training set of the current iteration self-adaptive depth forest model into the current iteration self-adaptive depth forest model, and outputting the class probability corresponding to each sample in the training set.
And calculating the classification precision of the self-adaptive depth forest model after the training set is trained by the current iterative self-adaptive depth forest model by using a training set classification precision calculation formula.
Step 5, judging whether the classification precision of the current iterative adaptive depth forest model is less than 95%; if yes, go to step 6, otherwise go to step 12.
And 6, obtaining the weight of each tree in the current iterative adaptive depth forest model.
(6.1) obtaining the weight of each tree in each random forest of the current iteration self-adaptive depth forest model by using the following self-defined quadratic convex optimization expression:
ξi k≥0,i∈K
wherein the content of the first and second substances,expressing xi when minimizing an objective functioni,wkI denotes the index of the sample in the training set, K denotes the total number of samples in the training set, Σ denotes the summation operation, ξi kRepresenting the basic loss of the ith sample in the training set in the kth random forest in the current iterative adaptive depth forest model, lambda representing a regularization coefficient, | | | | - | representing 2 norm operation, wkRepresenting the total weight of the kth forest in the adaptive depth forest model of the current iteration, s.t. representing a constraint condition symbol, T representing a subscript of a tree in the kth random forest in the adaptive depth forest model of the current iteration, and TkRepresenting the total number of trees, P, in the kth random forest in the adaptive depth forest model for the current iterationi (t,k)Representing a base of an ith sample in a t-th tree of a kth random forest in the adaptive depth forest model of the current iterationThis loss, w(t,k)Representing the weight of the t number of the kth random forest in the current iteration of the adaptive depth forest model, and representing the classification interval by tau;
and the basic loss expression of the ith sample in the t tree of the kth random forest in the adaptive depth forest model of the current iteration is as follows:
wherein Z isiIndicating an indicator variable, and if the label of the ith sample in the training set is 0, ZiIf the label of the ith sample in the training set is 1, then Zi1 denotes a multiplication operation, pi,0 (t,k)Represents the probability, p, of the positive sample in the t-th tree of the kth random forest in the adaptive depth forest model of the current iteration for the ith sample in the training seti,1 (t,k)And representing the probability of a negative sample in the t tree of the kth random forest in the adaptive depth forest model of the current iteration for the ith sample in the training set.
(6.2) combining the weight of each tree in each random forest of the current iteration self-adaptive depth forest model into the weight of the current iteration self-adaptive depth forest model;
and 7, obtaining the enhancement characteristics of the current iterative adaptive depth forest model on the test set.
And respectively obtaining the enhancement features of each sample in the test set, namely a positive sample and a negative sample, according to the weight of each tree in the current iteration adaptive depth forest model and through the enhancement feature solving formula of the current iteration adaptive depth forest model.
And forming the enhancement features of the adaptive depth forest model of the current iteration by the enhancement features of the positive samples and the negative samples of all the samples in the test set.
The enhanced feature solving formula of the current iteration self-adaptive depth forest model is as follows:
wherein v isi,0Represents the enhanced probability that the ith sample in the training set is a positive sample, pi,0 (t,k)Represents the probability of the positive sample in the t tree of the kth random forest in the adaptive depth forest model of the current iteration, vi,1Indicates the enhanced probability that the ith sample in the training set is a negative sample, pi,1 (t,k)And representing the probability of a negative sample in the t tree of the kth random forest in the adaptive depth forest model of the current iteration for the ith sample in the training set.
And 8, calculating the accuracy of the test set.
And calculating the accuracy of the test set by using the accuracy calculation method of the adaptive depth forest model test set of the current iteration according to the enhancement characteristics of the adaptive depth forest model of the current iteration on the test set.
The method for calculating the accuracy of the current iteration self-adaptive depth forest model test set comprises the following steps:
in the first step, for each sample in the test set, the enhanced feature value of the positive sample is greater than that of the negative sample, and the label of the sample is 0, then the sample is classified correctly, otherwise the sample is classified incorrectly.
And secondly, for each sample in the test set, if the enhanced characteristic value of the positive sample is smaller than that of the negative sample and the label of the sample is 1, the sample is classified correctly, otherwise, the sample is classified incorrectly.
And thirdly, counting the total number of correctly classified samples in the test set.
And fourthly, dividing the total number of the correctly classified samples in the test set by the total number of the samples in the test set to obtain the test set accuracy of the self-adaptive deep forest model.
And 9, judging whether the accuracy of the current iterative adaptive depth forest model test set is greater than the maximum accuracy of the test set, if so, executing the step 10, and otherwise, executing the step 12.
The maximum accuracy of the test set refers to the maximum value of the accuracy of the test set on the adaptive depth forest model with the depth of 1, 2, …, h.
And step 10, obtaining the enhancement characteristics of the current iteration self-adaptive depth forest model on a training set.
And respectively obtaining the enhancement features of each sample in the training set, namely a positive sample and a negative sample, through the enhancement feature solving formula of the current iterative adaptive depth forest model according to the weight of each tree in the current iterative adaptive depth forest model.
And forming the enhanced features of the adaptive depth forest model of the current iteration by the enhanced features of the positive samples and the negative samples of all the samples in the training set.
And 11, expanding the self-adaptive depth forest model of the current iteration.
And splicing the training set of the current iteration self-adaptive depth forest model and the enhanced features of the current iteration self-adaptive depth forest model end to obtain the training set of the current iteration self-adaptive depth forest model.
And taking the depth t +1 of the adaptive depth forest model of the current iteration as the depth t of the adaptive depth forest model in the current iteration.
And 12, judging whether the depth t of the current iterative adaptive depth forest model is greater than 100, if so, executing a step 13 after the trained adaptive depth forest model is obtained, and otherwise, executing a step 4.
And step 13, classifying all gait energy maps in the test set.
And inputting the test set of the current iterative adaptive depth forest model into the trained adaptive depth forest model, and classifying the gait energy map in the test set.
The effect of the present invention is further explained by combining the simulation experiment as follows:
1. simulation experiment conditions are as follows:
the hardware platform of the simulation experiment of the invention is as follows: intel Core i7-8700K @3.70GHz CPU, 64GB RAM, NVIDIA Geforce GTX1080 Ti GPU.
The software platform of the simulation experiment of the invention is as follows: python2.7 and TensorFlow1.4.
The data used in the simulation experiment of the invention is a public gait recognition data set Dataset B, which comprises contour silhouette images of human gait of 124 target persons under three walking conditions (common conditions, wearing overcoat and carrying package conditions) and 11 visual angles, the method synthesizes the gait energy images of 124 target persons, selects six gait energy images of each target person under the conditions of normal walking and 90-degree visual angle, and totally 744 gait energy images form a sample set, and randomly samples the sample set by a random combination sampling splicing method to form a training set and a test set.
2. Simulation content and result analysis:
in the simulation experiment of the invention, firstly, gait energy graphs of each target person in a database of the Dataset B under different conditions are synthesized according to the method of the invention, six gait energy graphs of each target person under normal walking conditions and with a visual angle of 90 degrees are selected, 744 gait energy graphs form a sample set, and the sample set is randomly sampled by using a random combination sampling splicing method to form a training set and a testing set. And training the adaptive depth forest model by using the training set, and then testing the trained adaptive depth forest model by using the test sample to obtain the classification precision of the test set.
The simulation experiment of the invention is to classify the gait energy chart of the test set constructed by the invention by applying the method of the invention and the Random forest (Mach. learning, 2001, Leo Breiman) of the prior art under the same data set.
In order to verify the classification effect of the method and the random forest in the prior art, the classification precision of the method and the random forest is respectively calculated and compared by using the following formula.
Accuracy of classification-correctly classified sample in test set/total number of samples in test set
TABLE 1 Classification accuracy List of the two methods
Simulation experiment method Random forest Method for producing a composite material
Accuracy of classification 75.43% 92.93%
The results of the classification accuracy calculation of the invention and the existing random forest technology are respectively listed in table 1, and as can be seen from table 1, the classification accuracy of the invention is 92.93%, which is much higher than that of the random forest classification method, and the classification accuracy of the gait energy map higher than that of the random forest method can be obtained.

Claims (8)

1. A classification method based on a self-adaptive depth forest human gait energy map is characterized in that the weight of each tree in a self-adaptive depth forest model is solved, the enhancement features of the self-adaptive depth forest model are obtained, the complexity of the depth forest model is determined according to the depth self-adaptation of the depth forest model, and the enhancement features of the obtained depth forest model are used for training the depth forest model; the method comprises the following specific steps:
(1) constructing a sample set:
(1a) carrying out background modeling and background real-time updating on 124 input target character videos to obtain a human body outline foreground image of each target character;
(1b) extracting a human body gait image from the human body outline foreground image of each target person video;
(1c) adding the gray values of all the human body contour images of each target person obtained in each gait cycle, and dividing the sum by the total number of the human body contour images of each target person to obtain a gait energy map corresponding to the gait cycle;
(1d) taking six gait energy maps of each target person under a normal walking condition and with a visual angle of 90 degrees, and forming a sample set by 744 gait energy maps;
(2) obtaining a training set and a test set:
(2a) randomly sampling a sample set by using a random combination sampling splicing method to obtain 992 positive samples and negative samples, and forming a training set by using all the positive samples and the negative samples;
(2b) randomly sampling a sample set by using a random combination sampling splicing method to obtain 992 positive samples and negative samples, and forming a test set by using all the positive samples and the negative samples;
(3) constructing an adaptive depth forest model:
(3a) building a self-adaptive depth forest model consisting of two random forests and two completely random forests;
(3b) setting the number of trees in two random forests and two complete random forests in the self-adaptive depth forest model to be 400, setting the maximum depth of the trees to be 8, setting bootstrap sample marks to be True, setting the number of parallel-processed cores to be the maximum number of cores of a server, and setting the depth of the self-adaptive depth forest model to be h;
(4) calculating the classification precision of the current iteration self-adaptive depth forest model:
(4a) inputting a training set of the current iteration self-adaptive depth forest model into the current iteration self-adaptive depth forest model, and outputting class probability corresponding to each sample in the training set;
(4b) calculating the classification precision of the depth forest model after the training set is trained by the current iterative adaptive depth forest model by using a training set classification precision calculation formula;
(5) judging whether the classification precision of the current iteration self-adaptive depth forest model is less than 95%; if yes, executing the step (6), otherwise, executing the step (12);
(6) obtaining the weight of each tree in the current iteration of the adaptive depth forest model:
(6a) obtaining the weight of each tree in each random forest of the current iterative self-adaptive depth forest model by using the following self-defined quadratic convex optimization expression:
ξi k≥0,i∈K
wherein the content of the first and second substances,expressing xi when minimizing an objective functioni,wkI denotes the index of the sample in the training set, K denotes the total number of samples in the training set, Σ denotes the summation operation, ξi kRepresenting the basic loss of the ith sample in the training set in the kth random forest in the current iterative adaptive depth forest model, lambda representing a regularization coefficient, | | | | - | representing 2 norm operation, wkRepresenting the total weight of the kth forest in the adaptive depth forest model of the current iteration, s.t. representing a constraint condition symbol, T representing a subscript of a tree in the kth random forest in the adaptive depth forest model of the current iteration, and TkRepresenting the total number of trees, P, in the kth random forest in the adaptive depth forest model for the current iterationi (t,k)Representing the fundamental loss, w, of the ith sample in the t tree of the kth random forest in the adaptive depth forest model of the current iteration(t,k)Representing the number of kth random forests in the adaptive depth forest model for the current iterationWeight, τ denotes the classification interval;
(6b) the weight of each tree in each random forest of the current iteration adaptive depth forest model is formed into the weight of the current iteration adaptive depth forest model;
(7) obtaining the enhancement characteristics of the current iteration self-adaptive depth forest model on the test set:
(7a) according to the weight of each tree in the current iteration self-adaptive depth forest model, respectively obtaining the enhancement features of each sample in the test set, namely a positive sample and a negative sample, through an enhancement feature solving formula of the current iteration self-adaptive depth forest model;
(7b) the enhancement features of the positive samples and the negative samples of all samples in the test set form the enhancement features of the adaptive depth model of the current iteration;
(8) and (3) calculating the accuracy of the test set:
calculating the accuracy of the test set by using the accuracy calculation method of the adaptive depth forest model test set of the current iteration according to the enhancement characteristics of the adaptive depth forest model of the current iteration on the test set;
(9) judging whether the accuracy of the current iterative adaptive depth forest model test set is greater than the maximum accuracy of the test set, if so, executing the step (10), otherwise, executing the step (12);
(10) obtaining the enhancement characteristics of the current iteration self-adaptive depth forest model on a training set:
(10a) respectively obtaining the enhancement features of each sample in the training set, namely a positive sample and a negative sample, through an enhancement feature solving formula of the current iterative adaptive depth forest model according to the weight of each tree in the current iterative adaptive depth forest model;
(10b) the enhancement features of the current iteration self-adaptive depth forest model are formed by the enhancement features of the positive samples and the negative samples of all samples in the training set;
(11) expanding the self-adaptive depth forest model of the current iteration:
(11a) splicing the training set of the current iteration self-adaptive depth forest model and the enhancement features of the current iteration self-adaptive depth forest model end to obtain the training set of the current iteration self-adaptive depth forest model;
(11b) taking the depth h +1 of the adaptive depth forest model of the current iteration as the depth h of the adaptive depth forest model during the current iteration;
(12) judging whether the depth h of the current iterative adaptive depth forest model is greater than 100, if so, executing the step (13) after obtaining the trained depth forest model, otherwise, executing the step (4);
(13) classifying all gait energy maps in the test set:
and inputting the test set of the current iterative self-adaptive depth forest model into the trained depth forest model, and classifying the gait energy map in the test set.
2. The classification method based on the adaptive deep forest human gait energy map as claimed in claim 1, characterized in that: the background modeling and background real-time updating in the step (1a) means that a previous frame and a background frame of a character target video are subtracted to obtain a human body outline foreground image of the target character video; and calling a function in a visual library to remove holes and scattered white points existing in the human body contour foreground image.
3. The classification method based on the adaptive deep forest human gait energy map as claimed in claim 1, characterized in that: the step of extracting the human body gait image in the step (1b) is that the distance between the leftmost point and the rightmost point of each human body contour is taken as the width of the human body contour, and the distance between the uppermost point and the lowermost point is taken as the height of the human body contour; the gait cycle is divided by the width change signal of the human body contour by utilizing the characteristic that the contour width of the human body is synchronously and periodically changed along with time; and (3) scratching out the human body outline image according to the width and the height of the human body outline, keeping the height-width ratio unchanged, and scaling all the human body outline images to the same size.
4. The classification method based on the adaptive deep forest human gait energy map as claimed in claim 1, characterized in that: the random combination sampling splicing method in the step (2) comprises the following steps:
the method comprises the following steps that firstly, 6 gait energy maps of an unselected human target are randomly selected from 124 human targets in a sample set to serve as a current processing object;
secondly, randomly selecting four pairs of gait energy maps from the current processing object, transversely splicing each pair of gait energy maps to form a positive sample, setting the label of the positive sample to be 0, and setting the number of the positive samples of each target character to be 4;
step three, randomly selecting a gait energy map from the current processing object;
fourthly, respectively and randomly selecting 4 different target characters from the sample set except the current target character, and respectively and randomly selecting one gait energy image from 6 gait energy images of the selected target character, wherein 4 gait energy images are total;
fifthly, pairing every two of the gait energy diagram obtained in the third step with the 4 state energy diagrams obtained in the fourth step to form 4 pairs of gait energy diagrams, transversely splicing the gait energy diagrams of each pair to form a negative sample, setting the label of the negative sample to be 1, wherein the number of the negative samples of each character target is 4;
and a sixth step: marking the current processing object as processed;
the seventh step: if 6 gait energy maps of each target person are marked as processed for 124 target persons, 496 positive samples and 496 negative samples are obtained in total, namely 992 samples in total, and if 6 gait energy maps of the target persons exist for 124 target persons, the first step is executed.
5. The classification method based on the adaptive deep forest human gait energy map as claimed in claim 1, characterized in that: in the step (6a), the basic loss expression of the ith sample in the t tree of the kth random forest in the adaptive depth forest model of the current iteration is as follows:
wherein Z isiIndicating an indicator variable, and if the label of the ith sample in the training set is 0, ZiIf the label of the ith sample in the training set is 1, then Zi1 denotes a multiplication operation, pi,0 (t,k)Represents the probability, p, of the positive sample in the t-th tree of the kth random forest in the adaptive depth forest model of the current iteration for the ith sample in the training seti,1 (t,k)And representing the probability of a negative sample in the t tree of the kth random forest in the adaptive depth forest model of the current iteration for the ith sample in the training set.
6. The classification method based on the adaptive deep forest human gait energy map as claimed in claim 5, characterized in that: in the step (7a), the enhanced feature solving formula of the current iteration adaptive depth forest model is as follows:
wherein v isi,0Indicates the enhanced probability that the ith sample in the training set is a positive sample, vi,1Indicating the enhanced probability that the ith sample in the training set is a negative sample.
7. The classification method based on the adaptive deep forest human gait energy map as claimed in claim 1, characterized in that: the method for calculating the accuracy of the current iteration self-adaptive depth forest model test set in the step (8) comprises the following steps:
step one, for each sample in a test set, if the enhancement characteristic value of a positive sample is greater than that of a negative sample and the label of the sample is 0, the sample is classified correctly, otherwise, the sample is classified incorrectly;
secondly, for each sample in the test set, if the enhancement characteristic value of the positive sample is smaller than that of the negative sample and the label of the sample is 1, the sample is classified correctly, otherwise, the sample is classified incorrectly;
thirdly, counting the total number of correctly classified samples in the test set;
and fourthly, dividing the total number of the correctly classified samples in the test set by the total number of the samples in the test set to obtain the test set accuracy of the deep forest model.
8. The classification method based on the adaptive deep forest human gait energy map as claimed in claim 1, characterized in that: the maximum accuracy of the test set in the step (9) is the maximum value of the accuracy of the test set on the adaptive depth forest model with the depth of 1, 2, …, h.
CN201811222012.3A 2018-10-19 2018-10-19 Classification method based on self-adaptive deep forest human gait energy map Active CN109255339B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201811222012.3A CN109255339B (en) 2018-10-19 2018-10-19 Classification method based on self-adaptive deep forest human gait energy map

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201811222012.3A CN109255339B (en) 2018-10-19 2018-10-19 Classification method based on self-adaptive deep forest human gait energy map

Publications (2)

Publication Number Publication Date
CN109255339A CN109255339A (en) 2019-01-22
CN109255339B true CN109255339B (en) 2021-04-06

Family

ID=65045465

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201811222012.3A Active CN109255339B (en) 2018-10-19 2018-10-19 Classification method based on self-adaptive deep forest human gait energy map

Country Status (1)

Country Link
CN (1) CN109255339B (en)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110706738B (en) * 2019-10-30 2020-11-20 腾讯科技(深圳)有限公司 Method, device, equipment and storage medium for predicting structure information of protein

Citations (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2001039655A2 (en) * 1999-12-06 2001-06-07 Trustees Of Boston University In-shoe remote telemetry gait analysis system
CN101241546A (en) * 2007-11-20 2008-08-13 西安电子科技大学 Method for compensating for gait binary value distortion
EP2439492A1 (en) * 2010-10-07 2012-04-11 Honeywell International, Inc. System and method for wavelet-based gait classification
CN103473539A (en) * 2013-09-23 2013-12-25 智慧城市系统服务(中国)有限公司 Gait recognition method and device
CN104200200A (en) * 2014-08-28 2014-12-10 公安部第三研究所 System and method for realizing gait recognition by virtue of fusion of depth information and gray-scale information
CN104299012A (en) * 2014-10-28 2015-01-21 中国科学院自动化研究所 Gait recognition method based on deep learning
CN105574510A (en) * 2015-12-18 2016-05-11 北京邮电大学 Gait identification method and device
CN107212890A (en) * 2017-05-27 2017-09-29 中南大学 A kind of motion identification and fatigue detection method and system based on gait information
US9811720B2 (en) * 2013-10-22 2017-11-07 Bae Systems Information And Electronic Systems Integration Inc. Mobile device based gait biometrics

Patent Citations (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2001039655A2 (en) * 1999-12-06 2001-06-07 Trustees Of Boston University In-shoe remote telemetry gait analysis system
CN101241546A (en) * 2007-11-20 2008-08-13 西安电子科技大学 Method for compensating for gait binary value distortion
EP2439492A1 (en) * 2010-10-07 2012-04-11 Honeywell International, Inc. System and method for wavelet-based gait classification
CN103473539A (en) * 2013-09-23 2013-12-25 智慧城市系统服务(中国)有限公司 Gait recognition method and device
US9811720B2 (en) * 2013-10-22 2017-11-07 Bae Systems Information And Electronic Systems Integration Inc. Mobile device based gait biometrics
CN104200200A (en) * 2014-08-28 2014-12-10 公安部第三研究所 System and method for realizing gait recognition by virtue of fusion of depth information and gray-scale information
CN104299012A (en) * 2014-10-28 2015-01-21 中国科学院自动化研究所 Gait recognition method based on deep learning
CN105574510A (en) * 2015-12-18 2016-05-11 北京邮电大学 Gait identification method and device
CN107212890A (en) * 2017-05-27 2017-09-29 中南大学 A kind of motion identification and fatigue detection method and system based on gait information

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
Deep Forest: Towards An Alternative to Deep Neural Networks;Zhi-Hua Zhou 等;《arXiv:1702.08835v2》;arXiv;20170531;1-10 *

Also Published As

Publication number Publication date
CN109255339A (en) 2019-01-22

Similar Documents

Publication Publication Date Title
CN106326886B (en) Finger vein image quality appraisal procedure based on convolutional neural networks
CN103955702B (en) SAR image terrain classification method based on depth RBF network
CN105956560B (en) A kind of model recognizing method based on the multiple dimensioned depth convolution feature of pondization
CN104392463B (en) Image salient region detection method based on joint sparse multi-scale fusion
CN107122375B (en) Image subject identification method based on image features
CN105809198B (en) SAR image target recognition method based on depth confidence network
CN103942568B (en) A kind of sorting technique based on unsupervised feature selection
CN108717568B (en) A kind of image characteristics extraction and training method based on Three dimensional convolution neural network
CN108510012A (en) A kind of target rapid detection method based on Analysis On Multi-scale Features figure
CN107784320B (en) Method for identifying radar one-dimensional range profile target based on convolution support vector machine
CN105931253B (en) A kind of image partition method being combined based on semi-supervised learning
CN108564129B (en) Trajectory data classification method based on generation countermeasure network
CN104268593A (en) Multiple-sparse-representation face recognition method for solving small sample size problem
CN103440471B (en) The Human bodys' response method represented based on low-rank
CN106682569A (en) Fast traffic signboard recognition method based on convolution neural network
CN105205449A (en) Sign language recognition method based on deep learning
CN104459668A (en) Radar target recognition method based on deep learning network
CN105243139A (en) Deep learning based three-dimensional model retrieval method and retrieval device thereof
CN108447057A (en) SAR image change detection based on conspicuousness and depth convolutional network
CN103714340A (en) Self-adaptation feature extracting method based on image partitioning
CN109255339B (en) Classification method based on self-adaptive deep forest human gait energy map
CN110472652A (en) A small amount of sample classification method based on semanteme guidance
CN112733665A (en) Face recognition method and system based on lightweight network structure design
CN103617417B (en) Automatic plant identification method and system
CN110633745A (en) Image classification training method and device based on artificial intelligence and storage medium

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