CN107748892B - Human behavior data segmentation method based on Mahalanobis distance - Google Patents

Human behavior data segmentation method based on Mahalanobis distance Download PDF

Info

Publication number
CN107748892B
CN107748892B CN201710875138.XA CN201710875138A CN107748892B CN 107748892 B CN107748892 B CN 107748892B CN 201710875138 A CN201710875138 A CN 201710875138A CN 107748892 B CN107748892 B CN 107748892B
Authority
CN
China
Prior art keywords
action
actions
basic
similarity
time period
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
CN201710875138.XA
Other languages
Chinese (zh)
Other versions
CN107748892A (en
Inventor
李军怀
田玲
王怀军
于蕾
王侃
安洋
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.)
Xian University of Technology
Original Assignee
Xian University of Technology
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 Xian University of Technology filed Critical Xian University of Technology
Priority to CN201710875138.XA priority Critical patent/CN107748892B/en
Publication of CN107748892A publication Critical patent/CN107748892A/en
Application granted granted Critical
Publication of CN107748892B publication Critical patent/CN107748892B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/22—Matching criteria, e.g. proximity measures
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/08—Feature extraction
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/12—Classification; Matching

Landscapes

  • Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Theoretical Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Image Analysis (AREA)
  • Measurement Of The Respiration, Hearing Ability, Form, And Blood Characteristics Of Living Organisms (AREA)

Abstract

The invention discloses a behavior data segmentation method based on Mahalanobis distance, which comprises the following steps: step 1, selecting a sliding window with the size of h to segment sensor data, calculating the Mahalanobis distance between window data, and dividing each window into a time interval action set according to a criterion; step 2, after the step 1 is finished, finding out all basic actions on the basis of time interval action division, and processing the hidden basic actions by adopting similarity analysis; and 3, after all basic actions are determined through the step 2, distinguishing transitional actions and disturbance processes according to whether two adjacent basic actions are the same, and finally determining all the basic actions and the transitional actions. The behavior data segmentation method based on the Mahalanobis distance divides the behavior of a person into two types of basic actions and transitional actions, and realizes accurate segmentation of basic action and transitional action data in a multi-behavior mode process.

Description

Human behavior data segmentation method based on Mahalanobis distance
Technical Field
The invention belongs to the technical field of human behavior recognition, and particularly relates to a human behavior data segmentation method based on Mahalanobis distance.
Background
Human behavior activity is a continuous process, human behavior data acquired from a sensor network is a continuous data stream, and behavior identification by reasonably dividing sample data of the continuous process into data segments according to what rule is a key problem to be solved by data segmentation. Data segmentation is a process of dividing a continuous process into several small time segments, each time segment is called a window, in other words, data segmentation is a process of discretizing sensor data at continuous time points. The behavior recognition takes a window as a basic processing unit, and the final recognition effect is influenced by too little or too much behavior data contained in the window, so that data segmentation is a very important link in the behavior recognition.
In conventional behavior recognition research, sliding window is one of the most common data segmentation techniques. Although the sliding window model has a good recognition effect in single independent action recognition, the method has difficulty in accurately segmenting the boundary of a basic action and a transition action between adjacent basic actions in the process of a multi-action mode.
Disclosure of Invention
The invention aims to provide a human body behavior data segmentation method based on Mahalanobis distance, which divides behavior actions of a human body into two types of basic actions and transitional actions and realizes accurate segmentation of basic action and transitional action data in a multi-behavior mode process.
The technical scheme adopted by the invention is that a human behavior data segmentation method based on Mahalanobis distance is implemented according to the following steps:
step 1, selecting a sliding window with the size of h to segment sensor data, calculating the Mahalanobis distance between window data, and dividing each window into a time interval action set according to a criterion;
step 2, after the step 1 is finished, finding out all basic actions on the basis of time interval action division, and processing the hidden basic actions by adopting similarity analysis;
and 3, after all basic actions are determined through the step 2, distinguishing transitional actions and disturbance processes according to whether two adjacent basic actions are the same, and finally determining all the basic actions and the transitional actions.
The invention is also characterized in that:
the step 1 is implemented according to the following steps:
step 1.1, a sliding window with the length of h is adopted to divide a two-dimensional data matrix S into k sections along the sampling direction, namely each window contains h sample data;
step 1.2, after the step 1.1, calculating the similarity among the action segments by using the Mahalanobis distance to form a similarity matrix;
step 1.3: repeating the step 1.2 until all action segments are divided into an action category u (i), i is more than or equal to 1 and less than or equal to k;
step 1.4, after the step 1.3, the categories of the action fragments are expanded along a time axis, and the action fragments which are continuous in time and belong to the same action category are divided into action sets in the same time interval
Figure BDA0001417955820000021
In step 2:
selecting the maximum similarity in a similarity matrix by using a behavior unit, and if the maximum similarity is greater than a threshold value alpha, regarding two action segments of a row and a column where the similarity exists as the same type of action; otherwise, the action fragment of the row is treated as a new type of action.
The step 2 is implemented according to the following steps:
step 2.1, searching a section with the longest running time in all time interval actions between the basic actions A and B, and taking the section as a reference time interval of a middle new basic action; if the time interval action with the longest running time is more than one, selecting the time interval action closest to the middle position;
step 2.2, after the step 2.1, comparing the similarity of the actions of the two adjacent time periods at the left and right of the reference time period with the reference time period, if the similarity of the actions of the two adjacent time periods is equal to the similarity of the reference time period, combining the actions of the two time periods with the reference time period as a new reference time period, and otherwise, selecting the action of the time period with high similarity and combining the action of the time period with the reference time period into the reference time period to form a new reference time period;
step 2.3, repeating the step 2.2 until a termination condition is met, considering that the hidden basic action is found out, and otherwise, continuing to perform the next iteration;
step 2.4, after the step 2.3 is completed, calculating the similarity between the newly determined basic action and all known basic actions, if the similarity is greater than a threshold value alpha, defining the basic action as the corresponding basic action, otherwise, indicating that the basic action is a new action which does not appear;
and 2.5, repeating the steps 2.1-2.4 until all basic actions are found, wherein all the basic actions in the multi-behavior mode are determined at the moment, and the action length of the time interval between the basic actions is less than the shortest running time of the basic actions.
The termination conditions in step 2.3 are specifically as follows:
the highest value of the similarity between actions of two adjacent time periods at the left and the right of the reference time period in the iteration and the reference time period is less than a threshold value alpha
Figure BDA0001417955820000041
Length of time of
Figure BDA0001417955820000042
Minimum run time T greater than basic actionmin。
Step 3 is specifically implemented according to the following method:
two adjacent basic actions are respectively two different actions A and B, and all sub-periods between the two basic actions are considered to belong to transition actions from the basic actions A to the basic actions B;
and if two adjacent basic actions are the same action A, all the sub-periods between the two basic actions are defined as the disturbance process of the basic action A.
The invention has the beneficial effects that:
the invention relates to a human body behavior data segmentation method based on Mahalanobis distance, which is used for dividing human behavior actions into two types of basic actions and transitional actions aiming at multi-mode identification problems of periodicity of actions of the same type, alternation of actions of different types, transitional actions during action switching and the like in a behavior identification process, and segmenting behavior data based on Mahalanobis distance, so that accurate segmentation of the basic action data and the transitional action data in a multi-behavior mode process is realized; and finally, on the basis, the data segmentation is carried out on the basic action by adopting the traditional sliding window technology and the basic action is used as the input data of the subsequent identification, so that the human behavior action is accurately identified.
Drawings
FIG. 1 is a schematic diagram of window segmentation and time-segment action segmentation in a human body behavior data segmentation method based on Mahalanobis distance according to the present invention;
FIG. 2 is a schematic diagram of data segmentation of basic actions and transitional actions in the method for segmenting human behavior data based on Mahalanobis distance.
Detailed Description
The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.
The invention relates to a human behavior data segmentation method based on Mahalanobis distance, which is implemented according to the following steps:
step 1, selecting a sliding window with the size of h to segment sensor data, calculating the Mahalanobis distance between window data, dividing each window into a time interval action set according to a criterion, and specifically implementing the following steps:
step 1.1, selecting a sliding window with a length of h to segment the sensor sample data matrix S along the time direction, and then representing the segmented sample data matrix as S ═ S1,S2,,Sk]T；
Step 1.2: after the step 1.1, calculating the similarity among the action segments by using the Mahalanobis distance, wherein the larger the Mahalanobis distance is, the smaller the similarity is, the larger the similarity is, and the smaller the Mahalanobis distance is, the greater the similarity is, and the action segments with the Mahalanobis distance smaller than a threshold value alpha are regarded as a type of action;
setting action segment wiAnd action fragment wjHas a Mahalanobis distance of mdijThen, the similarity matrix MD between all motion segments is as follows:
Figure BDA0001417955820000051
in the matrix MD, MD for the ith rowijJ is more than or equal to 1 and less than or equal to k, and the minimum min (md) is searchedij) If min (md)ij) Less than threshold α, then move segment wiAnd wjThe actions are regarded as the same kind of actions; otherwise, the action fragment wiTreating the action as a new type;
step 1.3: repeating the step 1.2 until all action segments are divided into an action category u (i), i is more than or equal to 1 and less than or equal to k;
step 1.4, after the step 1.3, the categories of the action fragments are expanded along a time axis, and the action fragments which are continuous in time and belong to the same action category are divided into action sets in the same time interval
Figure BDA0001417955820000061
As shown in fig. 1, the whole behavior process is initially divided into p time segment actions, the time segment actions are orderly arranged according to the time direction, and the p-th time segment action is represented as the following form:
Figure BDA0001417955820000062
Figure BDA0001417955820000063
belonging to the c-th action class, i.e.
Figure BDA0001417955820000064
Where C is 1,2, C, the p-th sub-period time length is noted
Figure BDA0001417955820000065
Step 2, after the step 1 is finished, finding out all basic actions on the basis of time interval action division, and processing the hidden basic actions by adopting similarity analysis;
defining the shortest running time of basic actions according to empirical knowledge, defining the time interval actions with the time interval action length larger than the shortest running time of the basic actions as the basic actions, and defining the basic actions belonging to the same action class as the same basic actions; if the sum of the lengths of all the time interval actions between two adjacent basic actions is larger than the shortest running time of the basic actions, the basic actions are still contained between the two basic actions, and the phenomenon is called as a basic action hidden phenomenon.
Setting two basic actions A and B which are determined to exist now, wherein the basic action A and the basic action B are adjacent, and the sum of the lengths of all time interval actions between the basic action A and the basic action B is larger than the shortest running time of the basic action;
the method is implemented according to the following steps:
step 2.1, searching a section with the longest running time in all time interval actions between the basic action A and the basic action B, and taking the section as a reference time interval of a middle new basic action; if the time interval action with the longest running time is more than one, selecting the time interval action closest to the middle position;
step 2.2, after the step 2.1, comparing the similarity of the actions of the two adjacent time periods at the left and right of the reference time period with the reference time period, if the similarity of the actions of the two adjacent time periods is equal to the similarity of the reference time period, combining the actions of the two time periods with the reference time period as a new reference time period, and otherwise, selecting the action of the time period with high similarity and combining the action of the time period with the reference time period into the reference time period to form a new reference time period;
step 2.3, repeating the step 2.2 until a termination condition is met, considering that the hidden basic action is found out, and otherwise, continuing to perform the next iteration;
the termination conditions were as follows:
the highest value of the similarity between actions of two adjacent time periods at the left and the right of the reference time period in the iteration and the reference time period is less than a threshold value alpha
Figure BDA0001417955820000071
Length of time of
Figure BDA0001417955820000072
Minimum run time T greater than basic actionmin；
Step 2.4, after the step 2.3 is completed, calculating the similarity between the newly determined basic action and all known basic actions, if the similarity is greater than a threshold value alpha, defining the basic action as the corresponding basic action, otherwise, indicating that the basic action is a new action which does not appear;
and 2.5, repeating the steps 2.1-2.4 until all basic actions are found, wherein all the basic actions in the multi-behavior mode are determined at the moment, and the action length of the time interval between the basic actions is less than the shortest running time of the basic actions.
Step 3, after determining all basic actions through step 2, distinguishing transition actions and disturbance processes according to whether two adjacent basic actions are the same, and finally determining all basic actions and transition actions, as shown in fig. 2:
two adjacent basic actions are respectively two different actions A and B, and all sub-periods between the two basic actions are considered to belong to transition actions from the basic actions A to the basic actions B;
and if two adjacent basic actions are the same action A, all the sub-periods between the two basic actions are defined as the disturbance process of the basic action A.
The invention relates to a human body behavior data segmentation method based on Mahalanobis distance, which is used for dividing human behavior actions into two types of basic actions and transitional actions aiming at multi-mode identification problems of periodicity of actions of the same type, alternation of actions of different types, transitional actions during action switching and the like in a behavior identification process, and segmenting behavior data based on Mahalanobis distance, so that accurate segmentation of the basic action data and the transitional action data in a multi-behavior mode process is realized. The data segmentation method is used in the human behavior recognition process, so that the accuracy of behavior recognition is improved to a great extent. Meanwhile, the data segmentation method also provides a new idea and method for other similar multi-modal recognition problems.

Claims (2)

1. A human behavior data segmentation method based on Mahalanobis distance is characterized by comprising the following steps:
step 1, selecting a sliding window with the size of h to segment sensor data, calculating the Mahalanobis distance between window data, and dividing each window into a time interval action set according to a criterion, wherein the method specifically comprises the following steps:
step 1.1, a sliding window with the length of h is adopted to divide a two-dimensional data matrix S into k sections along the sampling direction, namely each window contains h sample data;
step 1.2, after the step 1.1, calculating the similarity among the action segments by using the Mahalanobis distance to form a similarity matrix;
step 1.3: repeating the step 1.2 until all action segments are divided into an action category u (i), i is more than or equal to 1 and less than or equal to k;
step 1.4, after the step 1.3, the categories of the action fragments are expanded along a time axis, and the action fragments which are continuous in time and belong to the same action category are divided into action sets in the same time interval
Figure FDA0002633475950000011
Step 2, after the step 1 is completed, finding out all basic actions on the basis of time interval action division, and processing the hidden basic actions by adopting similarity analysis, wherein the method specifically comprises the following steps of:
the step 2 is specifically implemented according to the following steps:
step 2.1, searching a section with the longest running time in all time interval actions between the basic actions A and B, and taking the section as a reference time interval of a middle new basic action; if the time interval action with the longest running time is more than one, selecting the time interval action closest to the middle position;
step 2.2, after the step 2.1, comparing the similarity of the actions of the two adjacent time periods at the left and right of the reference time period with the reference time period, if the similarity of the actions of the two adjacent time periods is equal to the similarity of the reference time period, combining the actions of the two time periods with the reference time period as a new reference time period, and otherwise, selecting the action of the time period with high similarity and combining the action of the time period with the reference time period into the reference time period to form a new reference time period;
step 2.3, repeating the step 2.2 until a termination condition is met, considering that the hidden basic action is found out, and otherwise, continuing to perform the next iteration;
the termination conditions were specifically as follows: the highest value of the similarity between actions of two adjacent time periods at the left and the right of the reference time period in the iteration and the reference time period is less than a threshold value alpha
Figure FDA0002633475950000021
Length of time of
Figure FDA0002633475950000022
Minimum run time T greater than basic actionmin；
Step 2.4, after the step 2.3 is completed, calculating the similarity between the newly determined basic action and all known basic actions, if the similarity is greater than a threshold value alpha, defining the basic action as the corresponding basic action, otherwise, indicating that the basic action is a new action which does not appear;
step 2.5, repeating the step 2.1 to the step 2.4 until all basic actions are found out, wherein all the basic actions in the multi-behavior mode are determined at the moment, and the action length of the time interval between the basic actions is smaller than the shortest running time of the basic actions;
step 3, after all basic actions are determined through the step 2, distinguishing transition actions and disturbance processes according to whether two adjacent basic actions are the same, and finally determining all the basic actions and the transition actions; the method specifically comprises the following steps:
two adjacent basic actions are respectively two different actions A and B, and all sub-periods between the two basic actions are considered to belong to transition actions from the basic actions A to the basic actions B;
and if two adjacent basic actions are the same action A, all the sub-periods between the two basic actions are defined as the disturbance process of the basic action A.
2. The method for segmenting human behavior data based on Mahalanobis distance as claimed in claim 1, wherein in the step 2:
selecting the maximum similarity in a similarity matrix by using a behavior unit, and if the maximum similarity is greater than a threshold value alpha, regarding two action segments of a row and a column where the similarity exists as the same type of action; otherwise, the action fragment of the row is treated as a new type of action.
CN201710875138.XA 2017-09-25 2017-09-25 Human behavior data segmentation method based on Mahalanobis distance Active CN107748892B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201710875138.XA CN107748892B (en) 2017-09-25 2017-09-25 Human behavior data segmentation method based on Mahalanobis distance

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201710875138.XA CN107748892B (en) 2017-09-25 2017-09-25 Human behavior data segmentation method based on Mahalanobis distance

Publications (2)

Publication Number Publication Date
CN107748892A CN107748892A (en) 2018-03-02
CN107748892B true CN107748892B (en) 2020-11-17

Family

ID=61254732

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201710875138.XA Active CN107748892B (en) 2017-09-25 2017-09-25 Human behavior data segmentation method based on Mahalanobis distance

Country Status (1)

Country Link
CN (1) CN107748892B (en)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109034179B (en) * 2018-05-30 2022-03-22 河南理工大学 Rock stratum classification method based on Mahalanobis distance IDTW

Family Cites Families (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102298649B (en) * 2011-10-09 2012-11-28 南京大学 Space trajectory retrieval method of body movement data
CN102819549B (en) * 2012-04-16 2016-03-30 大连大学 Based on the human motion sequences segmentation method of Least-squares estimator characteristic curve
CN103310233B (en) * 2013-06-28 2016-03-23 青岛科技大学 With similarity method for digging between class behavior multi views and Activity recognition method
CN104317681B (en) * 2014-09-02 2017-09-08 上海交通大学 For the behavioral abnormal automatic detection method and detecting system of computer system
CN105631596B (en) * 2015-12-29 2020-12-29 山东鲁能软件技术有限公司 Equipment fault diagnosis method based on multi-dimensional piecewise fitting
CN106021926B (en) * 2016-05-20 2019-06-18 北京九艺同兴科技有限公司 A kind of real-time estimating method of human action sequence
CN106730771B (en) * 2017-01-05 2018-12-21 大连理工大学 A kind of basketball action data processing method divided based on unit movement

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
"A Self-updating Template Approach for Human Action Recognition Using Improved DTW";Changhui Li等;《Journal of Information and Computational Science》;20141101;第13卷(第11期);第4507-4515页 *

Also Published As

Publication number Publication date
CN107748892A (en) 2018-03-02

Similar Documents

Publication Publication Date Title
CN110610168B (en) Electroencephalogram emotion recognition method based on attention mechanism
CN109003625B (en) Speech emotion recognition method and system based on ternary loss
Xu et al. Activity auto-completion: Predicting human activities from partial videos
CN109271876B (en) Video motion detection method based on time evolution modeling and multi-example learning
CN108845982B (en) Chinese word segmentation method based on word association characteristics
CN108804453A (en) A kind of video and audio recognition methods and device
CN109034062B (en) Weak supervision abnormal behavior detection method based on time sequence consistency
CN103577841A (en) Human body behavior identification method adopting non-supervision multiple-view feature selection
Stern et al. Most discriminating segment–Longest common subsequence (MDSLCS) algorithm for dynamic hand gesture classification
CN111860278B (en) Human behavior recognition algorithm based on deep learning
Gu et al. An advanced deep learning approach for safety helmet wearing detection
CN106203255A (en) A kind of pedestrian based on time unifying heavily recognition methods and system
CN105389471A (en) Method for reducing training set of machine learning
CN107748892B (en) Human behavior data segmentation method based on Mahalanobis distance
CN111881263A (en) Service recommendation online optimization method for intelligent home scene
CN108764026B (en) Video behavior detection method based on time sequence detection unit pre-screening
CN112418269B (en) Social media network event propagation key time prediction method, system and medium
CN110032585B (en) Time sequence double-layer symbolization method and device
Sun et al. Latent structured perceptrons for large-scale learning with hidden information
CN108509990A (en) A kind of sequential key assignments type industrial process data Parallel analytic method
Raghuraman et al. Efficient thyroid disease prediction and comparative study using machine learning algorithms
Jia et al. Deep learning with convolutional neural networks for sleep arousal detection
CN111914724A (en) Continuous Chinese sign language identification method and system based on sliding window segmentation
Gharacheh et al. Detection of metamorphic malware based on hmm: A hierarchical approach
Dai et al. A novel method for simultaneous gesture segmentation and recognition based on HMM

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