CN103235928A - Gait recognition method with monitoring mechanism - Google Patents
Gait recognition method with monitoring mechanism Download PDFInfo
- Publication number
- CN103235928A CN103235928A CN2013100047130A CN201310004713A CN103235928A CN 103235928 A CN103235928 A CN 103235928A CN 2013100047130 A CN2013100047130 A CN 2013100047130A CN 201310004713 A CN201310004713 A CN 201310004713A CN 103235928 A CN103235928 A CN 103235928A
- Authority
- CN
- China
- Prior art keywords
- gait
- model
- information
- state
- layer
- 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.)
- Pending
Links
Images
Landscapes
- Image Analysis (AREA)
- Measurement Of The Respiration, Hearing Ability, Form, And Blood Characteristics Of Living Organisms (AREA)
Abstract
Disclosed is a gait recognition method with a monitoring mechanism. The gait recognition method is characterized in that a three-layer dynamic Bayesian network model is adopted, a first layer of the model is featured by the contour of gait of a human body, and a state of the first layer at a current moment is only relevant to a state of the first layer at a next moment; a second layer of the model is featured by frame difference images of the gait, and a state of the second layer at the current moment is relevant to a state of the second layer at the next moment and is relevant to the state of the first layer at the current moment and the state of the first layer at the next moment; and a third layer of the model is a monitoring layer and is relevant to the state of the second layer at the current moment and the state of the first layer at the current moment. The gait recognition method has the advantages that the traditional model structure is improved, and the gait recognition method is additionally provided with the monitoring mechanism; optimization learning for parameters is obviously improved, and the gait recognition rate is greatly increased; and the gait recognition method has an actual application significance and enormous social benefit.
Description
Technical field
The present invention relates to computer science and technical field of image processing, a kind of gait recognition method with supervision mechanism is provided especially.
Background technology
As long-range biological characteristic authentication technology, Gait Recognition more and more is subject to people's attention. and Gait Recognition is exactly to carry out remote authentication according to the posture that the people walks. and gait has non-infringement, is difficult to camouflage, unlike fingerprint or iris recognition when the feature extraction need be identified target and keep close contact, so gait is the biological characteristic of the potentialization in long-distance video monitoring field.
Increasing to the research of Gait Recognition in the last few years, mostly be to identify behind the gait profile by extraction people walking, technology can be divided into two classes; The one, utilize the static information of gait profile, as Kim Gait Recognition [Kim D based on active contour model and motion prediction has been proposed, Paik J. Gait recognition using active shape model and motion prediction [J], Computer Vision IET, 2010,4 (1): 25 – 36.], mainly be to utilize the profile information of human body walking to adopt active shape model (ASM) to identify, this class algorithm relies on the static profile of human body, usually worn, knapsack influences recognition effect, as shown in Figure 1.The 2nd, adopt the multidate information that extracts profile, the algorithm that this class is studied is a lot, as the gait recognition method [Wang Kejun based on gait energygram picture (GEI) and 2 dimension principal component analysis (PCA)s, Liu Lili, beautifully adorned Xian is firelight or sunlight. based on the gait recognition method [J] of gait energygram picture and 2 dimension principal component analysis (PCA)s, China's image graphics journal, 2009,14 (12): 2503-2509], the method utilizes the GEI image as the gait feature image, carry out 2 dimension principal component analysis (PCA)s, but same because GEI image be the body gait frame and the image that constitutes of mean value, must be subjected to overcoat, influences such as knapsack are normally, overcoat, in the unitary class training sample of knapsack higher discrimination is arranged, at overcoat, knapsack, discrimination is lower under the normal blending mode, as shown in Figure 2; Fig. 2 for same individual at the GEI image of normally walking, wearing under overcoat, the knapsack, the appearance profile static information difference of GEI image under three kinds of states is very big as can be seen, so discrimination is lower; In order to improve the too many multidate information of GEI missing image, the method based on active energy figure (AEI) and two-dimentional partial projection that Zhang proposes is identified [Zhang E H, Zhao Y W, Xiong W. Active energy image plus 2DLPP for gait recognition [J], Signal Processing, 2010,90 (7): 2295-2302], wherein the AEI image is to be made of the frame difference image stack, dynamic gait feature in the time of can well reacting human motion, the recognition methods novelty, but the AEI image has been ignored the static information of human body simultaneously, and as shown in Figure 3, the AEI image is fine to dynamic information representation as can be seen, leg exercise information particularly, but exist overcoat equally, knapsack is to the influence of profile static information; The mode that the incomplete gait profile to extracting that static information Chen when walking for keeping gait proposes is taked to set up frame difference energygram (FDEI) makes up gait feature, set up hidden Markov model (HMM) simultaneously and explain [Chen C H, Liang J M, Zhao H. Frame difference energy image for gait recognition with incomplete silhouettes [J], Pattern Recognition Letters, 2009,30 (11): 977 – 984.], recognition effect is good, wherein frame difference energygram (FDEI) is to superpose with the frame-to-frame differences image, and the partial summation of motionless maintenance static state constitutes the FDEI image during with the walking of human body, and the FDEI image can well be expressed static state and the multidate information of human body; As shown in Figure 4, frame difference image can be good at the expressive movement feature, but since with the static part of gait walking as static information, so that the FDEI image that constitutes is influenced by knapsack still is bigger.Chen has also proposed based on the Gait Recognition of the dynamic bayesian network of bilayer (DBN) [Zhang Erhu, Zhao Yongwei. the Gait Recognition [J] of utilizing dynamic position to change, China's image graphics journal, 2009,14 (9): 1756-1763.], at first divide some sections whole gait sequence, ground floor is expressed with dynamic texture, the second layer is expressed with hidden Markov model, and the method novelty has taken into full account the gait temporal characteristics.The Gait Recognition based on arm and leg exercise that Faezeh proposes, analyze the motion of shank and arm and identify gait [Tafazzoli F, Safabakhsh R. Model-based human gait recognition using leg and arm movements [J], Engineering Applications of Artificial Intelligence, 2010,23 (8): 1237-1246.]; Only use the recognition methods of the multidate information of shank that more application is also arranged, recognition effect and robustness are also better, as shown in Figure 5; The motion dynamic part of Fig. 5 for from the GEI image, extracting, image is 8 gray level images, movement threshold is taken as 220, and image is the image under three kinds of conditions of normal walking, overcoat, knapsack of same individual, and same individual's shank gait information characteristics is very similar as can be seen.
Gait Recognition is mostly set about from the characteristic image of setting up gait on the recognition methods step, utilizes the characteristic image of gait to carry out Feature Extraction identification.Recognition methods be divided into method based on characteristic matching [Zhao Yongwei, Zhang Erhu, Lu Jiwen. the Gait Recognition [J] of many features and various visual angles information fusion, Chinese image graphics journal, 2009,14 (3): 387-393; Bashir K, Xiang T, Gong S G. Gait recognition without subject cooperation [J], Pattern Recognition Letters, 2010,31 (13): 2052-2060.], based on method [Chen C H, the Liang J M of dynamic bayesian network, Zhu X C. Gait recognition based on improved dynamic Bayesian networks [J], Pattern Recognition, 2011,44 (4): 988 – 995; Suk H, Sin B K, Lee S W. Hand gesture recognition based on dynamic Bayesian network framework [J], Pattern Recognition, 43 (9): 3059-3072.]; Wherein the method based on dynamic bayesian network is the focus of studying at present, the temporal characteristics of the gait that it takes into full account.And hidden Markov model (HMM) and improved model thereof have obtained using widely [Bae J as a kind of form of dynamic bayesian network, Tomizuka M. Gait phase analysis based on a Hidden Markov Model [J], Mechatronics, 2011,21 (6): 961-970; Cheng M H, Ho M F, Chung-Lin Huang. Gait analysis for human identification through manifold learning and HMM [J], Pattern Recognition, 2008,41 (8): 2541-2553.], as coupled hidden markov model (CHMM) [Waleed H, Kasabov N. Reduced feature-set based parallel CHMM speech recognition systems [J], Information Sciences, 2003,156 (1-2): 21-38.], embedded Markov model (EHMM), the Gait Recognition [Zhang Qianjin based on built-in type hidden Markov model (EHMM) as the Zhang Qianjin proposition, Xu Suli. based on the Gait Recognition [J] of built-in type hidden Markov model, information and control, 2010,39 (1): 25-29.], Gait Recognition [the Zhang Erhu that changes of the dynamic position of utilization that proposes of Zhang Erhu wherein, Zhao Yongwei. the Gait Recognition [J] of utilizing dynamic position to change, China's image graphics journal, 2009,14 (9): 1756-1763], method has incorporated the characteristic of sequential, use based on maximum entropy Markov model (MEMM) as the gait classification device, the discrimination height, the robustness of model is also better.
Study from gait recognition method, more and more researchers all is to set about from the feature of expressing gait, wish to find more essential gait feature, the model application of considering temporal aspect in the recognition methods is also more and more, wishes to reach the better recognition effect with time-based probability inference.
Therefore, people expect to obtain a kind of gait recognition method with supervision mechanism.
Summary of the invention:
The objective of the invention is to come the human body identity is authenticated by gait, a kind of gait recognition method with supervision mechanism is provided.
The invention provides a kind of gait recognition method with supervision mechanism, it is characterized in that: it adopts three layers dynamic bayesian network model to carry out Gait Recognition, wherein: it is feature that the model ground floor adopts the body gait profile, and the state of current time is only relevant with next one moment state; It is feature that the model second layer adopts the gait frame difference image, and the state of current time is relevant constantly with next, and relevant with ground floor current time state and next moment state; Model is monitor layer for the 3rd layer, and is relevant with second layer current time state and ground floor current time state.
In the described gait recognition method with supervision mechanism, the gait frame difference image feature in the model second layer adopts the method for setting up proper vector to express, and following steps are adopted in the foundation of proper vector:
At first cut apart the human body frame difference image with the rectangle frame of fixed size, the human body frame difference image is carried out staging treating;
Next calculates the moment of inertia of every section rectangle frame;
At last, come the construction feature vector with the value of moment of inertia.
Described gait recognition method with supervision mechanism also satisfies following requirement: monitor layer is according to the ground floor state of current time, judge namely whether the human body contour outline image is the abnormal behaviour of non-body gait, decides current model whether to carry out model learning and identification; Simultaneously, also according to current time second layer state, namely whether the judgment frame difference image is the abnormal behaviour of non-body gait, decides current model whether to carry out model learning and identification.
Described Bayesian network model with supervision mechanism contains 3 state variables
,
,
And 5 observational variables
Wherein
Be used for expressing the static frames information of gait sequence, namely
t-1 moment gait frame,
tMoment gait frame, above-mentioned gait frame information only comprises the information of the gait static state of current time;
Be used for expressing frame difference image,
It is right to be used for representing
With
The supervision of state; Observational variable
,
,
Be respectively the joint angle of gait static frames information, highly, width information; Observational variable
,
Be respectively speed and the amplitude information of the frame difference of gait motion;
In 3 layers dynamic bayesian network model, multidate information and the static information of gait when the state of model ground floor and the second layer is walked in order to describe the people; Dynamic probability process hypothesis is Ma Shi in each model layer, and namely following probability constantly is only relevant with current time and irrelevant constantly with the past:
Depend on
Because gait walking is the process of a sequential, reflection be with constantly
tThe variation of the attitude of gait frame, amplitude, profile and rhythm; With constantly
tThe gait amplitude when walking that the multidate information that changes can reflect the people, the variation of rhythm, and with constantly
tThe static information that changes can well be expressed the attitude of gait, the information of profile profile; Multidate information is subjected to the expression that is used for of current time and previous moment static information in model.
In the described gait recognition method with supervision mechanism, gait frame difference image feature adopts the method for setting up proper vector to express, and following steps are adopted in the foundation of proper vector:
Cut apart the human body frame difference image with the rectangle frame of fixed size at first from top to bottom, the human body frame difference image is divided into some sections; Next calculates the moment of inertia of every section rectangle frame; At last, come the construction feature vector with the value of moment of inertia; That is:
Described gait recognition method with supervision mechanism, model is monitor layer for the 3rd layer, monitor layer judges namely according to current time ground floor state whether the human body contour outline image is the abnormal behaviour of non-body gait, decides current model whether to carry out model learning and identification; And according to current time second layer state, namely whether the judgment frame difference image is the abnormal behaviour of non-body gait, decides current model whether to carry out model learning and identification;
The gait walking is the process of a sequential, at the unit interval sheet
tIn, not only include the appearance profile static information of being expressed by the gait sequence frame, the rhythmic multidate information when also including human body walking such as the motion amplitude of being expressed by frame difference image, speed; In the expression of multidate information, when the people walks in gait, be usually expressed as the alternating rhythmical swing of left and right sides limbs, when this swing is spent view from people's body side surface 90, show as human body about the part by behind the forward direction, again by after forward motion process, in this process gait frame difference: the forward direction part of frame difference and frame difference back expressed about human motion dynamic motion characteristic partly to part;
The Bayesian network model with supervision mechanism that constructs by as above multidate information and static information: the Bayesian network model with supervision mechanism contains 3 state variables
,
,
And 5 observational variables
Wherein
Be used for expressing the static frames information of gait sequence, namely as Fig. 6
t-1 moment gait frame,
tMoment gait frame, this gait frame information has only comprised the information of the gait static state of current time, as appearance profile, attitude etc.;
Be used for expressing frame difference image,
It is right to be used for representing
With
The supervision of state; Observational variable
,
,
For the joint angle of gait static frames information, highly, width information; Observational variable
,
Speed and amplitude information for the frame difference of gait motion;
Gait recognition method with supervision mechanism adopts 3 layers dynamic bayesian network model, multidate information and the static information of gait when the state of ground floor and the second layer is used for describing the people and walks in the model, the 3rd layer is used as monitor layer, and dynamic probability process hypothesis is Ma Shi (Markovian) in every layer, and namely following probability constantly is only relevant with current time and irrelevant constantly with the past:
Depend on
Because gait walking is the process of a sequential, reflection be with constantly
tThe variation of the attitude of gait frame, amplitude, profile and rhythm; With constantly
tThe gait amplitude when walking that the multidate information that changes can well reflect the people, the variation of rhythm, and with constantly
tThe static information that changes can well be expressed the attitude of gait, the information of profile profile; Multidate information is subjected to the expression that is used for of current time and previous moment static information in model, the sound attitude information of the fine fusion gait walking of model energy;
Reasoning dynamic bayesian network model calculates exactly at given observation sequence
Calculate latent state variable
Marginal probability
, by calculating the joint probability distribution of all state nodes, and then marginalisation, and then calculate the probability distribution of all state nodes, the overall joint probability distribution of model reasoning is:
In following formula, the joint probability distribution of arbitrary state node is:
Model learning comes the parameter of estimation model to carry out according to given training data, order
Expression
tState constantly,
The expression status switch,
Expression
tObservation data constantly,
The expression observation sequence.The task of model learning is to come the parameter of estimation model according to given training data, for given training observation sequence
, by model parameter
The maximum likelihood method estimation, namely
,
Observation is incomplete, so pass through Expectation-maximization(EM) algorithm carries out iterative:
(14)
Wherein
Expression the
nParameter estimation during inferior iteration by following formula iteration convergence to a local extremum, can reach local optimum at least;
In view of the Gait Recognition that uses a model, identification is a reasoning iterative process based on dynamic bayesian network, and is given
RThe individual model that trains
, the corresponding people's of each model gait wherein, by test, observation sequence is
, then determine classification by following formula:
Wherein
Be model
Prior probability, be taken as mean value 1/
R, establish
Model parameter is
, then
, observation sequence is given, then
, then following formula can be derived as:
Determine classification by above formula, and then identification.
Among the present invention, regard gait walking the process of a sequential as, at the unit interval sheet
tIn, not only include the appearance profile static information of being expressed by the gait sequence frame, the rhythmic multidate information when also including human body walking such as the motion amplitude of being expressed by frame difference image, speed.In the expression of multidate information, when the people walks in gait, be usually expressed as the alternating rhythmical swing of left and right sides limbs, when this swing is spent view from people's body side surface 90, show as human body about the part by behind the forward direction, again by after forward motion process, in this process gait frame difference: the forward direction part of frame difference and frame difference back expressed to part about human motion partly the dynamic motion characteristic as shown in Figure 6.
The Bayesian network model that has supervision mechanism by constructing of as above multidate information and static information is as shown in Figure 7: the Bayesian network model with supervision mechanism contains 3 state variables
,
,
And 5 observational variables
Wherein
Be used for expressing the static frames information of gait sequence, namely as Fig. 6
t-1 moment gait frame,
tMoment gait frame, this gait frame information has only comprised the information of the gait static state of current time, as appearance profile, attitude etc.;
Be used for expressing frame difference image, as shown in Figure 6,
It is right to be used for representing
With
The supervision of state; Observational variable
,
,
For the joint angle of gait static frames information, highly, width information; Observational variable
,
Speed and amplitude information for the frame difference of gait motion.
Gait recognition method with supervision mechanism adopts 3 layers dynamic bayesian network model, multidate information and the static information of gait when the state of ground floor and the second layer is used for describing the people and walks in the model, the 3rd layer is used as monitor layer, and dynamic probability process hypothesis is Ma Shi (Markovian) in every layer, and namely following probability constantly is only relevant with current time and irrelevant constantly with the past:
Depend on
Because gait walking is the process of a sequential, reflection be with constantly
tThe variation of the attitude of gait frame, amplitude, profile and rhythm; With constantly
tThe gait amplitude when walking that the multidate information that changes can well reflect the people, the variation of rhythm, and with constantly
tThe static information that changes can well be expressed the attitude of gait, the information of profile profile.Multidate information is subjected to the expression that is used for of current time and previous moment static information in model, the sound attitude information of the fine fusion gait walking of model energy.
Model has following advantage: 1) can effective fusion sound attitude information during human body walking; 2) details that body part changes in the time of can catching human motion namely adopts frame difference image to express gait information; 3) compare with traditional HMM, model reduces parameter space greatly; 4) model has supervision mechanism, can effectively carry out parameter learning and identification.
The sound attitude information of the fine fusion gait walking of model energy, give full expression to the temporal characteristics of gait, compare with other models, the model that the present invention proposes is different with embedded Markov model EHMM, the latter adopts Embedded mode embedded many Markov chains in single Markov chain, form class is similar to the markov of layering, is state and the identification that comes learning model from single observation sequence, as the method for document 14 employings; And coupled hidden markov model CHMM is that the Markov chain by many symmetries constitutes, and is difficult to express the relation of different levels structure, and namely multidate information and static information are this according to belonging to relation, also do not have supervisory role simultaneously.
Description of drawings:
Fig. 1 is the static contour images of gait;
Fig. 2 for same individual's GEI image (from left to right respectively state be that a is normal, b overcoat, knapsack);
Fig. 3 be same individual AEI image (from left to right respectively state be that a is normal, b overcoat, knapsack);
Fig. 4 (first arranges respectively that state is: the normal static part of a, b overcoat static part, c knapsack static part from left to right for same individual's FDEI image; Second arrange respectively that state is from left to right: the normal FDEI figure of d, e overcoat FDEI figure, f knapsack FDEI figure);
Fig. 5 for motion parts information in the gait (from left to right respectively state be that a is normal, b overcoat, knapsack);
Fig. 6 is that gait multidate information image (being respectively from left to right: (a) t-1 moment gait frame, (b) t moment gait frame, (c)) forward frame difference figure, (d) back are to frame difference figure);
Fig. 7 is the dynamic bayesian network model with supervision mechanism;
Fig. 8 sets up proper vector for rectangle frame.
Embodiment:
Specify: because of in the art, the Gait Recognition associated picture is to show in groups and use by the mode described in " description of drawings " traditionally; Do not add fractionation, explanation hereby so express custom according to specialty.
A kind of gait recognition method with supervision mechanism, it adopts three layers dynamic bayesian network model to carry out Gait Recognition, and wherein: it is feature that the model ground floor adopts the body gait profile, and the state of current time is only relevant with next one moment state; It is feature that the model second layer adopts the gait frame difference image, and the state of current time is relevant constantly with next, and relevant with ground floor current time state and next moment state; Model is monitor layer for the 3rd layer, and ground floor current time state relevant with second layer current time state is correlated with.
In the described gait recognition method with supervision mechanism, the gait frame difference image feature in the model second layer adopts the method for setting up proper vector to express, and following steps are adopted in the foundation of proper vector:
At first cut apart the human body frame difference image with the rectangle frame of fixed size, the human body frame difference image is carried out staging treating;
Next calculates the moment of inertia of every section rectangle frame;
At last, come the construction feature vector with the value of moment of inertia.
Described gait recognition method with supervision mechanism also satisfies following requirement: monitor layer is according to the ground floor state of current time, judge namely whether the human body contour outline image is the abnormal behaviour of non-body gait, decides current model whether to carry out model learning and identification; Simultaneously, also according to current time second layer state, namely whether the judgment frame difference image is the abnormal behaviour of non-body gait, decides current model whether to carry out model learning and identification.
Described Bayesian network model with supervision mechanism contains 3 state variables
,
,
And 5 observational variables
Wherein
Be used for expressing the static frames information of gait sequence, namely as Fig. 6
t-1 moment gait frame,
tMoment gait frame, above-mentioned gait frame information only comprises the information of the gait static state of current time, as appearance profile, attitude etc.;
Be used for expressing frame difference image, as shown in Figure 6,
It is right to be used for representing
With
The supervision of state; Observational variable
,
,
For the joint angle of gait static frames information, highly, width information; Observational variable
,
Speed and amplitude information for the frame difference of gait motion;
In 3 layers dynamic bayesian network model, multidate information and the static information of gait when the state of model ground floor and the second layer is walked in order to describe the people; Dynamic probability process hypothesis is Ma Shi (Markovian) in each model layer, and namely following probability constantly is only relevant with current time and irrelevant constantly with the past:
Depend on
Because gait walking is the process of a sequential, reflection be with constantly
tThe variation of the attitude of gait frame, amplitude, profile and rhythm; With constantly
tThe gait amplitude when walking that the multidate information that changes can reflect the people, the variation of rhythm, and with constantly
tThe static information that changes can well be expressed the attitude of gait, the information of profile profile; Multidate information is subjected to the expression that is used for of current time and previous moment static information in model.
In the described gait recognition method with supervision mechanism, gait frame difference image feature adopts the method for setting up proper vector to express, and following steps are adopted in the foundation of proper vector:
Cut apart the human body frame difference image with the rectangle frame of fixed size at first from top to bottom, the human body frame difference image is divided into some sections; Next calculates the moment of inertia of every section rectangle frame; At last, come the construction feature vector with the value of moment of inertia; That is:
, wherein
mBe taken as the pixel value of pixel, the frame difference is bianry image, and white portion should be 255, but pixel value setting constant value is 1 in the test for the ease of calculating;
rBe the distance of pixel to rectangular centre, setting rectangular centre point is coordinate (0,0) point, namely
, in the rectangular area dynamically the pixel coordinate at position be (
X, y), namely the coordinate points of white portion in the image is set up proper vector with this eigenwert;
Described gait recognition method with supervision mechanism, model is monitor layer for the 3rd layer, monitor layer judges namely according to current time ground floor state whether the human body contour outline image is the abnormal behaviour of non-body gait, decides current model whether to carry out model learning and identification; And according to current time second layer state, namely whether the judgment frame difference image is the abnormal behaviour of non-body gait, decides current model whether to carry out model learning and identification;
The gait walking is the process of a sequential, at the unit interval sheet
tIn, not only include the appearance profile static information of being expressed by the gait sequence frame, the rhythmic multidate information when also including human body walking such as the motion amplitude of being expressed by frame difference image, speed; In the expression of multidate information, when the people walks in gait, be usually expressed as the alternating rhythmical swing of left and right sides limbs, when this swing is spent view from people's body side surface 90, show as human body about the part by behind the forward direction, again by after forward motion process, in this process gait frame difference: the forward direction part of frame difference and frame difference back expressed about human motion dynamic motion characteristic (as shown in Figure 6) partly to part;
The Bayesian network model with supervision mechanism that constructs by as above multidate information and static information: the Bayesian network model with supervision mechanism contains 3 state variables
,
,
And 5 observational variables
Wherein
Be used for expressing the static frames information of gait sequence, namely as Fig. 6
t-1 moment gait frame,
tMoment gait frame, this gait frame information has only comprised the information of the gait static state of current time, as appearance profile, attitude etc.;
Be used for expressing frame difference image, as shown in Figure 6,
It is right to be used for representing
With
The supervision of state; Observational variable
,
,
For the joint angle of gait static frames information, highly, width information; Observational variable
,
Speed and amplitude information for the frame difference of gait motion;
Gait recognition method with supervision mechanism adopts 3 layers dynamic bayesian network model, multidate information and the static information of gait when the state of ground floor and the second layer is used for describing the people and walks in the model, the 3rd layer is used as monitor layer, and dynamic probability process hypothesis is Ma Shi (Markovian) in every layer, and namely following probability constantly is only relevant with current time and irrelevant constantly with the past:
Depend on
Because gait walking is the process of a sequential, reflection be with constantly
tThe variation of the attitude of gait frame, amplitude, profile and rhythm; With constantly
tThe gait amplitude when walking that the multidate information that changes can well reflect the people, the variation of rhythm, and with constantly
tThe static information that changes can well be expressed the attitude of gait, the information of profile profile; Multidate information is subjected to the expression that is used for of current time and previous moment static information in model, the sound attitude information of the fine fusion gait walking of model energy;
Reasoning dynamic bayesian network model calculates exactly at given observation sequence
Calculate latent state variable
Marginal probability
, by calculating the joint probability distribution of all state nodes, and then marginalisation, and then calculate the probability distribution of all state nodes, the overall joint probability distribution of model reasoning is:
In following formula, the joint probability distribution of arbitrary state node is:
Conditional probability distribution is:
Model learning comes the parameter of estimation model to carry out according to given training data, order
Expression
tState constantly,
The expression status switch,
Expression
tObservation data constantly,
The expression observation sequence.The task of model learning is to come the parameter of estimation model according to given training data, for given training observation sequence
, by model parameter
The maximum likelihood method estimation, namely
,
Observation is incomplete, so pass through Expectation-maximization(EM) algorithm carries out iterative:
Wherein
Expression the
nParameter estimation during inferior iteration by following formula iteration convergence to a local extremum, can reach local optimum at least;
In view of the Gait Recognition that uses a model, identification is a reasoning iterative process based on dynamic bayesian network, and is given
RThe individual model that trains
, the corresponding people's of each model gait wherein, by test, observation sequence is
, then determine classification by following formula:
Wherein
Be model
Prior probability, be taken as mean value 1/
R, establish
Model parameter is
, then
, observation sequence is given, then
, then following formula can be derived as:
Determine classification by above formula, and then identification.
Claims (5)
1. gait recognition method with supervision mechanism, it is characterized in that: it adopts three layers dynamic bayesian network model to carry out Gait Recognition, wherein: it is feature that the model ground floor adopts the body gait profile, and the state of current time is only relevant with next one moment state; It is feature that the model second layer adopts the gait frame difference image, and the state of current time is relevant constantly with next, and relevant with ground floor current time state and next moment state; Model is monitor layer for the 3rd layer, and is relevant with second layer current time state and ground floor current time state.
2. according to the described a kind of gait recognition method with supervision mechanism of claim 1, it is characterized in that: the gait frame difference image feature in the model second layer adopts the method for setting up proper vector to express, and following steps are adopted in the foundation of proper vector:
At first, cut apart the human body frame difference image with the rectangle frame of fixed size, the human body frame difference image is carried out staging treating;
Secondly, calculate the moment of inertia of every section rectangle frame;
At last, come the construction feature vector with the value of moment of inertia.
3. according to the described a kind of gait recognition method with supervision mechanism of claim 2, it is characterized in that: described gait recognition method with supervision mechanism also satisfies following requirement: monitor layer is according to the ground floor state of current time, judge namely whether the human body contour outline image is the abnormal behaviour of non-body gait, decides current model whether to carry out model learning and identification; Simultaneously, also according to current time second layer state, namely whether the judgment frame difference image is the abnormal behaviour of non-body gait, decides current model whether to carry out model learning and identification.
4. according to the described a kind of gait recognition method with supervision mechanism of claim 3, it is characterized in that:
Described Bayesian network model with supervision mechanism contains 3 state variables
,
,
And 5 observational variables
Wherein
Be used for expressing the static frames information of gait sequence, namely
t-1 moment gait frame,
tMoment gait frame, above-mentioned gait frame information only comprises the information of the gait static state of current time;
Be used for expressing frame difference image,
It is right to be used for representing
With
The supervision of state; Observational variable
,
,
For the joint angle of gait static frames information, highly, width information; Observational variable
,
Speed and amplitude information for the frame difference of gait motion;
In 3 layers dynamic bayesian network model, multidate information and the static information of gait when the state of model ground floor and the second layer is walked in order to describe the people; Dynamic probability process hypothesis is Ma Shi in each model layer, and namely following probability constantly is only relevant with current time and irrelevant constantly with the past:
Depend on
Because gait walking is the process of a sequential, reflection be with constantly
tThe variation of the attitude of gait frame, amplitude, profile and rhythm; With constantly
tThe gait amplitude when walking that the multidate information that changes can reflect the people, the variation of rhythm, and with constantly
tThe static information that changes can well be expressed the attitude of gait, the information of profile profile; Multidate information is subjected to the expression that is used for of current time and previous moment static information in model.
5. according to claim 2,3 or 4 described a kind of gait recognition methods with supervision mechanism, it is characterized in that:
In the described gait recognition method with supervision mechanism, gait frame difference image feature adopts the method for setting up proper vector to express, and following steps are adopted in the foundation of proper vector:
Cut apart the human body frame difference image with the rectangle frame of fixed size at first from top to bottom, the human body frame difference image is divided into some sections; Next calculates the moment of inertia of every section rectangle frame; At last, come the construction feature vector with the value of moment of inertia; That is:
Described gait recognition method with supervision mechanism, model is monitor layer for the 3rd layer, monitor layer judges namely according to current time ground floor state whether the human body contour outline image is the abnormal behaviour of non-body gait, decides current model whether to carry out model learning and identification; And according to current time second layer state, namely whether the judgment frame difference image is the abnormal behaviour of non-body gait, decides current model whether to carry out model learning and identification;
The gait walking is the process of a sequential, at the unit interval sheet
tIn, not only include the appearance profile static information of being expressed by the gait sequence frame, the rhythmic multidate information when also including the human body walking of being expressed by frame difference image: motion amplitude, speed; In the expression of multidate information, when the people walks in gait, be usually expressed as the alternating rhythmical swing of left and right sides limbs, when this swing is spent view from people's body side surface 90, show as human body about the part by behind the forward direction, again by after forward motion process, in this process gait frame difference: the forward direction part of frame difference and frame difference back expressed about human motion dynamic motion characteristic partly to part;
The Bayesian network model with supervision mechanism that constructs by as above multidate information and static information: the Bayesian network model with supervision mechanism contains 3 state variables
,
,
And 5 observational variables
Wherein
Be used for expressing the static frames information of gait sequence, namely
t-1 moment gait frame,
tMoment gait frame, this gait frame information has only comprised the information of the gait static state of current time: appearance profile, attitude;
Be used for expressing frame difference image,
It is right to be used for representing
With
The supervision of state; Observational variable
,
,
For the joint angle of gait static frames information, highly, width information; Observational variable
,
Speed and amplitude information for the frame difference of gait motion;
Gait recognition method with supervision mechanism adopts 3 layers dynamic bayesian network model, multidate information and the static information of gait when the state of ground floor and the second layer is used for describing the people and walks in the model, the 3rd layer is used as monitor layer, and dynamic probability process hypothesis is Ma Shi (Markovian) in every layer, and namely following probability constantly is only relevant with current time and irrelevant constantly with the past:
Depend on
Because gait walking is the process of a sequential, reflection be with constantly
tThe variation of the attitude of gait frame, amplitude, profile and rhythm; With constantly
tThe gait amplitude when walking that the multidate information that changes can well reflect the people, the variation of rhythm, and with constantly
tThe static information that changes can well be expressed the attitude of gait, the information of profile profile; Multidate information is subjected to the expression that is used for of current time and previous moment static information in model, and model can merge the sound attitude information of gait walking;
Reasoning dynamic bayesian network model calculates exactly at given observation sequence
Calculate latent state variable
Marginal probability
, by calculating the joint probability distribution of all state nodes, and then marginalisation, and then calculate the probability distribution of all state nodes, the overall joint probability distribution of model reasoning is:
In following formula, the joint probability distribution of arbitrary state node is:
Model learning comes the parameter of estimation model to carry out according to given training data, order
Expression
tState constantly,
The expression status switch,
Expression
tObservation data constantly,
The expression observation sequence; The task of model learning is to come the parameter of estimation model according to given training data, for given training observation sequence
, by model parameter
The maximum likelihood method estimation, that is:
Observation is incomplete, so pass through Expectation-maximization(EM) algorithm carries out iterative:
Wherein
Expression the
nParameter estimation during inferior iteration by following formula iteration convergence to a local extremum, reaches local optimum at least;
In view of the Gait Recognition that uses a model, identification is a reasoning iterative process based on dynamic bayesian network, and is given
RThe individual model that trains
, the corresponding people's of each model gait wherein, by test, observation sequence is
, then determine classification by following formula:
Wherein
Be model
Prior probability, be taken as mean value 1/
R, establish
Model parameter is
, then
, observation sequence is given, then
, then following formula is derived as:
Determine classification by above formula, and then identify.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN2013100047130A CN103235928A (en) | 2013-01-08 | 2013-01-08 | Gait recognition method with monitoring mechanism |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN2013100047130A CN103235928A (en) | 2013-01-08 | 2013-01-08 | Gait recognition method with monitoring mechanism |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN103235928A true CN103235928A (en) | 2013-08-07 |
Family
ID=48883967
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN2013100047130A Pending CN103235928A (en) | 2013-01-08 | 2013-01-08 | Gait recognition method with monitoring mechanism |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN103235928A (en) |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107292231A (en) * | 2017-05-10 | 2017-10-24 | 广东技术师范学院 | A kind of method and system of the gesture recognition based under three-dimensional reconstruction |
| CN107864168A (en) * | 2016-09-22 | 2018-03-30 | 华为技术有限公司 | A kind of method and system of network data flow classification |
| WO2018223505A1 (en) * | 2017-06-06 | 2018-12-13 | 深圳市科迈爱康科技有限公司 | Gait-identifying wearable device |
| CN109002785A (en) * | 2018-07-05 | 2018-12-14 | 西安交通大学 | Gait recognition method based on movement timing energy diagram |
| CN111310587A (en) * | 2020-01-19 | 2020-06-19 | 中国计量大学 | Gait feature representation and feature extraction method based on fade-out motion trajectory graph |
| CN113407907A (en) * | 2021-06-04 | 2021-09-17 | 电子科技大学 | Hierarchical system structure function learning method fusing incomplete monitoring sequence |
| WO2021192682A1 (en) * | 2020-03-24 | 2021-09-30 | ソニーセミコンダクタソリューションズ株式会社 | Information processing device, information processing method, and program |
| CN113893517A (en) * | 2021-11-22 | 2022-01-07 | 动者科技(杭州)有限责任公司 | Rope skipping true and false judgment method and system based on difference frame method |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20040228503A1 (en) * | 2003-05-15 | 2004-11-18 | Microsoft Corporation | Video-based gait recognition |
| CN101226597A (en) * | 2007-01-18 | 2008-07-23 | 中国科学院自动化研究所 | Method and system for recognizing nights pedestrian based on thermal infrared gait |
-
2013
- 2013-01-08 CN CN2013100047130A patent/CN103235928A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20040228503A1 (en) * | 2003-05-15 | 2004-11-18 | Microsoft Corporation | Video-based gait recognition |
| CN101226597A (en) * | 2007-01-18 | 2008-07-23 | 中国科学院自动化研究所 | Method and system for recognizing nights pedestrian based on thermal infrared gait |
Non-Patent Citations (2)
| Title |
|---|
| 杨旗 等: "动静态信息融合及动态贝叶斯网络的步态识别", 《中国图象图形学报》 * |
| 赵悦 等: "一种基于半监督主动学习的动态贝叶斯网络算法", 《信息与控制》 * |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107864168A (en) * | 2016-09-22 | 2018-03-30 | 华为技术有限公司 | A kind of method and system of network data flow classification |
| US10999175B2 (en) | 2016-09-22 | 2021-05-04 | Huawei Technologies Co., Ltd. | Network data flow classification method and system |
| CN107292231A (en) * | 2017-05-10 | 2017-10-24 | 广东技术师范学院 | A kind of method and system of the gesture recognition based under three-dimensional reconstruction |
| WO2018223505A1 (en) * | 2017-06-06 | 2018-12-13 | 深圳市科迈爱康科技有限公司 | Gait-identifying wearable device |
| CN109002785A (en) * | 2018-07-05 | 2018-12-14 | 西安交通大学 | Gait recognition method based on movement timing energy diagram |
| CN111310587A (en) * | 2020-01-19 | 2020-06-19 | 中国计量大学 | Gait feature representation and feature extraction method based on fade-out motion trajectory graph |
| WO2021192682A1 (en) * | 2020-03-24 | 2021-09-30 | ソニーセミコンダクタソリューションズ株式会社 | Information processing device, information processing method, and program |
| CN113407907A (en) * | 2021-06-04 | 2021-09-17 | 电子科技大学 | Hierarchical system structure function learning method fusing incomplete monitoring sequence |
| CN113407907B (en) * | 2021-06-04 | 2022-04-12 | 电子科技大学 | Hierarchical system structure function learning method fusing incomplete monitoring sequence |
| CN113893517A (en) * | 2021-11-22 | 2022-01-07 | 动者科技(杭州)有限责任公司 | Rope skipping true and false judgment method and system based on difference frame method |
| CN113893517B (en) * | 2021-11-22 | 2022-06-17 | 动者科技(杭州)有限责任公司 | Rope skipping true and false judgment method and system based on difference frame method |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN103235928A (en) | Gait recognition method with monitoring mechanism | |
| Nadeem et al. | Human actions tracking and recognition based on body parts detection via Artificial neural network | |
| Jalal et al. | Shape and motion features approach for activity tracking and recognition from kinect video camera | |
| Arunnehru et al. | Human action recognition using 3D convolutional neural networks with 3D motion cuboids in surveillance videos | |
| CN105320944B (en) | A kind of human body behavior prediction method based on human skeleton motion information | |
| Kamal et al. | A hybrid feature extraction approach for human detection, tracking and activity recognition using depth sensors | |
| Kim et al. | Simultaneous gesture segmentation and recognition based on forward spotting accumulative HMMs | |
| Jalal et al. | Human daily activity recognition with joints plus body features representation using Kinect sensor | |
| CN101807245B (en) | Artificial neural network-based multi-source gait feature extraction and identification method | |
| CN107122736B (en) | Human body orientation prediction method and device based on deep learning | |
| CN114067358B (en) | Human body posture recognition method and system based on key point detection technology | |
| CN107256386A (en) | Human behavior analysis method based on deep learning | |
| WO2008007471A1 (en) | Walker tracking method and walker tracking device | |
| CN111160294B (en) | Gait recognition method based on graph convolution network | |
| CN106815578A (en) | A kind of gesture identification method based on Depth Motion figure Scale invariant features transform | |
| CN109902565A (en) | The Human bodys' response method of multiple features fusion | |
| CN102156994B (en) | Joint positioning method for single-view unmarked human motion tracking | |
| Krzeszowski et al. | Gait recognition based on marker-less 3D motion capture | |
| Sheu et al. | Improvement of human pose estimation and processing with the intensive feature consistency network | |
| Brubaker et al. | Video-based people tracking | |
| Liu et al. | Adaptive recognition method for VR image of Wushu decomposition based on feature extraction | |
| Hong et al. | Gait recognition using multi-bipolarized contour vector | |
| Huang et al. | View-independent behavior analysis | |
| Jin et al. | Research on Human Action Recognition Based on Global-Local Features of Video | |
| Guan et al. | Research on algorithm of human gait recognition based on sparse representation |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20130807 |
|
| RJ01 | Rejection of invention patent application after publication |