CN103489000A - Achieving method of human movement recognition training system - Google Patents
Achieving method of human movement recognition training system Download PDFInfo
- Publication number
- CN103489000A CN103489000A CN201310428277.XA CN201310428277A CN103489000A CN 103489000 A CN103489000 A CN 103489000A CN 201310428277 A CN201310428277 A CN 201310428277A CN 103489000 A CN103489000 A CN 103489000A
- Authority
- CN
- China
- Prior art keywords
- action
- tri
- vector
- frame
- characteristic
- 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)
Abstract
The invention discloses an achieving method of a human movement recognition training system. The method includes the following steps of (1) collecting training information, (2) normalizing sample data, (3) extracting a characteristic three-dimensional vector set, and (4) recognizing and identifying motions. According to the method, a reduced precision discretization coordinate algorithm is used for simplifying operation information representation and reducing calculation amount of a characteristic extraction algorithm, and meanwhile a multi-level feature matching algorithm is used for accelerating the recognition speed. The achieving method is simple and reasonable in logical design and algorithm, feasible, reliable and easy to achieve.
Description
Technical field
The invention belongs to the computer engineering design field, related to a kind of based on machine learning, implementation method produce the training system of particular person body action recognizer for length.
Background technology
Action recognition is very popular in recent years research field, by image-capturing apparatus, completes in the short period of time the identifying to human action, and is converted to the operational order of the equipment such as computing machine; Thereby be used as a kind of effective input medium and be applied to game, film making etc. widely in application.
The problem that at first action recognition will solve is to find the position of human action, and the position of human action is the foundation of action recognition, is commonly referred to as " concern position ".Because usually pay close attention to position behaviour face, the position that hand etc. are exposed, its color is with environment, clothes has larger difference, so can separate with non-concern lane place paying close attention to position color, for paying close attention to determining of position, the color histogram of general employing color-based distribution statistics is foundation, specifically, that human action is captured as to static image to be identified, according to image zones of different (two regional center positions, size have any one difference these two zones be zones of different) Color Statistical go out color histogram, then each, regional statistic histogram and default histogram compare, find the most similar zone as last concern position.
But this method requires very high to color, people dress clothes, surrounding environment and the colour of skin close with the colour of skin is close, single etc. the factor of surround lighting tone all can cause discrimination to decline to a great extent, and the method can only obtain which position of paid close attention to position at image, and can't obtain the implication that embodies of paying close attention to position.
For eliminating the impact of color on image recognition, existing recognition technology first is converted into gray-scale map to image to be identified usually, then the gray-scale map obtained is identified.After image to be identified is converted into gray-scale map, need to search out and pay close attention to position according to features such as lines of outline trend, each several part position relationships from integral body by the recognition system through the training of the artificial intelligence technologys such as neural network, for example from people's full-length picture, find face or hand.
At present, in the computer vision library item order of increasing income, adopted a kind of image recognition algorithm based on the simple feature cascade, adopt the action recognition process of this algorithm to be broadly divided into two parts, at first to pass through training process generating feature file, the image of catching is identified according to the tag file generated by identifying afterwards, obtained recognition result.This algorithm has been obtained effect preferably, but it still exists some shortcomings, as larger as the calculated amount of tag file, recognition speed is slow etc.Therefore, still need the image training recognizer that a kind of calculated amount is few, recognition speed is fast on market.
Summary of the invention
The object of the invention is to for the deficiencies in the prior art, provide a kind of precision discretize coordinate algorithm that falls to mean and reduce the calculated amount of feature extraction algorithm to simplify movable information, adopted the implementation method of multi-level features matching algorithm with the human action recognition training system of quickening recognition speed simultaneously.
To achieve these goals, the present invention has adopted following technical scheme:
A kind of implementation method of human action recognition training system comprises the following steps:
(1) collection of training information:
Gather the exercises sample that need to be identified by depth finding camera and third party SDK, action ID(Action ID:Int32), logarithm (Frame Count:Integer), depth information frame (Depth Frame #n:Integer Array) and the bone information frame (Skeleton Frame #n:Integer Array) of denomination of dive (Action Name:String), action message frame and, when the sample data by collecting is stored, its storage format comprises following field:;
(2) sample data Regularization:
A. unified coordinate system, carry out coordinate transform to depth information data and the bone information data of every frame, and they are unified in same world coordinate system;
B. discretize skeleton point coordinate, carry out the discretize processing to each bone information data point coordinate in present frame, and the step of discretize is:
I) the same skeleton point coordinate in adjacent each frame according to front and back, calculate the motion vector of this skeleton point;
Ii) search the unit cube in the target discrete coordinates system that this skeleton point is corresponding;
Iii) the motion vector of this skeleton point is synthesized in the current motion vector of unit cube;
C. discretize depth information data coordinates, at discretize skeleton point coordinate time, if this skeleton point is the hand point, near the depth data point this skeleton point is also carried out to discretize, the unit cube corresponding to the depth information data after discretize has the motion vector identical with this skeleton point;
D. each frame is repeated to above-mentioned a, b, c step, each action to be identified is generated to the sample tri-vector collection of a correspondence;
(3) extraction of characteristic 3 D vector set:
By the action ID in training sample data structure, know the corresponding human action of sample tri-vector collection after regularization, move for this all tri-vector collection that all training samples generate, all will move corresponding sample tri-vector collection as this; Sample tri-vector collection is carried out to feature extraction, obtain the characteristic 3 D vector set of this action, concrete extraction algorithm is as follows:
A. calculate each three-dimensional coordinate point and concentrate in whole tri-vectors the number of times occurred;
B. calculate its characteristic coefficient according to its occurrence number, that is, and the number of characteristic coefficient=three-dimensional coordinate point occurrence number/this action tri-vector collection;
If c. characteristic coefficient is greater than 50%, think that the tri-vector collection of this point belongs to the characteristic 3 D vector set;
(4) action recognition and identification:
The action to be identified for any one, obtain its tri-vector collection by step (1), (2), then adopt following algorithm to be identified:
A. identified fast
Calculate the distance of characteristic 3 D vectors all in this tri-vector collection and current system, if this distance is less than the corresponding threshold values of this characteristic 3 D vector, so using the action of this characteristic 3 D vector correspondence as a candidate actions;
B. meticulous identification
To all candidate actions, according to ascending sequence of distance, then successively to the sample tri-vector collection in each candidate actions, calculate the distance of the tri-vector collection of it and action to be identified, if wherein minor increment is less than default threshold values, system assert that action to be identified is current candidate actions automatically so.
In said method, further remark additionally, each element in described depth information frame is the depth information on its respective coordinates.
In said method, further remark additionally, described bone information frame is comprised of the skeleton point coordinate data.
In said method further supplementary notes, the logarithm of described action message frame refer in an action have how many to depth information frame and bone information frame.
SDK, the abbreviation of Software Development Kit, Chinese i.e. " SDK (Software Development Kit) ".Broadly refer to the set of relevant documentation, example and the instrument of a certain class software of auxiliary development.SDK is that some are used to specific software package, software frame, hardware platform, operating system etc. to create the set of the developing instrument of application software by the software engineer, and generally speaking SDK is the SDK that the application program under the developing Windows platform is used.It can simply provide some files of application programming interfaces API for certain programming language, but also may comprise can with the hardware of the complexity of certain embedded system communication.General instrument comprises the utility for debugging and other purposes.SDK also often comprises code sample, supportive technical notes or other the supporting documentation for basic reference clarification doubtful point.
The present invention is based on multiframe depth image data and the skeleton data that gathered by specific hardware, by the present invention for the action recognition field and custom-designed algorithm, realize that a human action trains and recognition system.
Advantage of the present invention:
1. in implementation method of the present invention, the precision discretize is fallen in raw data and process, the noise that not only can reduce like this raw data can also reduce operand, to accelerate recognition speed.
2. in implementation method of the present invention, adopt the multi-level features matching algorithm, further accelerated the action training recognition speed.
3. the logical design of implementation method of the present invention and algorithm advantages of simple, feasible reliable, easily realize.
The accompanying drawing explanation
Fig. 1 is sample data storage organization schematic diagram in the present invention.
Fig. 2 is coordinate discretize schematic diagram in the present invention.
Fig. 3 is action tri-vector collection schematic diagram in the present invention.
In Fig. 2, Fig. 3: the A-unit cube; Different high-precision dot in the B-original coordinates, after discretize, may be mapped in same unit cube; Tri-vector collection after C-action discretize.
Embodiment
Below in conjunction with embodiment and accompanying drawing, the present invention is further described.
Embodiment 1:
A kind of implementation method of human action recognition training system comprises the following steps:
(1) collection of training information:
Gather the exercises sample that need to be identified by depth finding camera and third party SDK, action ID(Action ID:Int32), logarithm (Frame Count:Integer), depth information frame (Depth Frame #n:Integer Array) and the bone information frame (Skeleton Frame #n:Integer Array) of denomination of dive (Action Name:String), action message frame and, when the sample data by collecting is stored, its storage format comprises following field:;
(2) sample data Regularization:
A. unified coordinate system, carry out coordinate transform to depth information data and the bone information data of every frame, and they are unified in same world coordinate system;
B. discretize skeleton point coordinate, carry out the discretize processing to each bone information data point coordinate in present frame, and the step of discretize is:
I) the same skeleton point coordinate in adjacent each frame according to front and back, calculate the motion vector of this skeleton point;
Ii) search the unit cube in the target discrete coordinates system that this skeleton point is corresponding;
Iii) the motion vector of this skeleton point is synthesized in the current motion vector of unit cube;
C. discretize depth information data coordinates, at discretize skeleton point coordinate time, if this skeleton point is the hand point, near the depth data point this skeleton point is also carried out to discretize, the unit cube corresponding to the depth information data after discretize has the motion vector identical with this skeleton point;
D. each frame is repeated to above-mentioned a, b, c step, each action to be identified is generated to the sample tri-vector collection of a correspondence;
(3) extraction of characteristic 3 D vector set:
By the action ID in training sample data structure, know the corresponding human action of sample tri-vector collection after regularization, move for this all tri-vector collection that all training samples generate, all will move corresponding sample tri-vector collection as this; Sample tri-vector collection is carried out to feature extraction, obtain the characteristic 3 D vector set of this action, concrete extraction algorithm is as follows:
A. calculate each three-dimensional coordinate point and concentrate in whole tri-vectors the number of times occurred;
B. calculate its characteristic coefficient according to its occurrence number, that is, and the number of characteristic coefficient=three-dimensional coordinate point occurrence number/this action tri-vector collection;
If c. characteristic coefficient is greater than 50%, think that the tri-vector collection of this point belongs to the characteristic 3 D vector set;
(4) action recognition and identification:
The action to be identified for any one, obtain its tri-vector collection by step (1), (2), then adopt following algorithm to be identified:
A. identified fast
Calculate the distance of characteristic 3 D vectors all in this tri-vector collection and current system, if this distance is less than the corresponding threshold values of this characteristic 3 D vector, so using the action of this characteristic 3 D vector correspondence as a candidate actions;
B. meticulous identification
To all candidate actions, according to ascending sequence of distance, then successively to the sample tri-vector collection in each candidate actions, calculate the distance of the tri-vector collection of it and action to be identified, if wherein minor increment is less than default threshold values, system assert that action to be identified is current candidate actions automatically so.
In said method, further remark additionally, each element in described depth information frame is the depth information on its respective coordinates.
In said method, further remark additionally, described bone information frame is comprised of the skeleton point coordinate data.
In said method further supplementary notes, the logarithm of described action message frame refer in an action have how many to depth information frame and bone information frame.
Claims (4)
1. the implementation method of a human action recognition training system, is characterized in that, this implementation method comprises the following steps:
(1) collection of training information:
Gather the exercises sample that need to be identified by depth finding camera and third party SDK, and, when the sample data by collecting is stored, its storage format comprises following field: logarithm, depth information frame and the bone information frame of action ID, denomination of dive, action message frame;
(2) sample data Regularization:
A. unified coordinate system, carry out coordinate transform to depth information data and the bone information data of every frame, and they are unified in same world coordinate system;
B. discretize skeleton point coordinate, carry out the discretize processing to each bone information data point coordinate in present frame, and the step of discretize is:
I) the same skeleton point coordinate in adjacent each frame according to front and back, calculate the motion vector of this skeleton point;
Ii) search the unit cube in the target discrete coordinates system that this skeleton point is corresponding;
Iii) the motion vector of this skeleton point is synthesized in the current motion vector of unit cube;
C. discretize depth information data coordinates, at discretize skeleton point coordinate time, if this skeleton point is the hand point, near the depth data point this skeleton point is also carried out to discretize, the unit cube corresponding to the depth information data after discretize has the motion vector identical with this skeleton point;
D. each frame is repeated to above-mentioned a, b, c step, each action to be identified is generated to the sample tri-vector collection of a correspondence;
(3) extraction of characteristic 3 D vector set:
By the action ID in training sample data structure, know the corresponding human action of sample tri-vector collection after regularization, move for this all tri-vector collection that all training samples generate, all will move corresponding sample tri-vector collection as this; Sample tri-vector collection is carried out to feature extraction, obtain the characteristic 3 D vector set of this action, concrete extraction algorithm is as follows:
A. calculate each three-dimensional coordinate point and concentrate in whole tri-vectors the number of times occurred;
B. calculate its characteristic coefficient according to its occurrence number, that is, and the number of characteristic coefficient=three-dimensional coordinate point occurrence number/this action tri-vector collection;
If c. characteristic coefficient is greater than 50%, think that the tri-vector collection of this point belongs to the characteristic 3 D vector set;
(4) action recognition and identification:
The action to be identified for any one, obtain its tri-vector collection by step (1), (2), then adopt following algorithm to be identified:
A. identified fast
Calculate the distance of characteristic 3 D vectors all in this tri-vector collection and current system, if this distance is less than the corresponding threshold values of this characteristic 3 D vector, so using the action of this characteristic 3 D vector correspondence as a candidate actions;
B. meticulous identification
To all candidate actions, according to ascending sequence of distance, then successively to the sample tri-vector collection in each candidate actions, calculate the distance of the tri-vector collection of it and action to be identified, if wherein minor increment is less than default threshold values, system assert that action to be identified is current candidate actions automatically so.
2. the implementation method of human action recognition training system according to claim 1, it is characterized in that: each element in described depth information frame is the depth information on its respective coordinates.
3. the implementation method of human action recognition training system according to claim 1, it is characterized in that: described bone information frame is comprised of the skeleton point coordinate data.
4. according to the implementation method of the arbitrary described human action recognition training system of claim 1-3, it is characterized in that: the logarithm of described action message frame refer in an action have how many to depth information frame and bone information frame.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201310428277.XA CN103489000A (en) | 2013-09-18 | 2013-09-18 | Achieving method of human movement recognition training system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201310428277.XA CN103489000A (en) | 2013-09-18 | 2013-09-18 | Achieving method of human movement recognition training system |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN103489000A true CN103489000A (en) | 2014-01-01 |
Family
ID=49829207
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201310428277.XA Pending CN103489000A (en) | 2013-09-18 | 2013-09-18 | Achieving method of human movement recognition training system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN103489000A (en) |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103942526A (en) * | 2014-01-17 | 2014-07-23 | 山东省科学院情报研究所 | Linear feature extraction method for discrete data point set |
| CN104091053A (en) * | 2014-06-26 | 2014-10-08 | 李南君 | Method and equipment for automatically detecting behavior pattern |
| CN104317386A (en) * | 2014-06-25 | 2015-01-28 | 西南科技大学 | Action recognition method of posture sequence finite-state machine |
| CN106203503A (en) * | 2016-07-08 | 2016-12-07 | 天津大学 | A kind of action identification method based on skeleton sequence |
| CN106203484A (en) * | 2016-06-29 | 2016-12-07 | 北京工业大学 | A kind of human motion state sorting technique based on classification layering |
| CN106650687A (en) * | 2016-12-30 | 2017-05-10 | 山东大学 | Posture correction method based on depth information and skeleton information |
| CN107225573A (en) * | 2017-07-05 | 2017-10-03 | 上海未来伙伴机器人有限公司 | The method of controlling operation and device of robot |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20100215257A1 (en) * | 2009-02-25 | 2010-08-26 | Honda Motor Co., Ltd. | Capturing and recognizing hand postures using inner distance shape contexts |
| CN102749993A (en) * | 2012-05-30 | 2012-10-24 | 无锡掌游天下科技有限公司 | Motion recognition method based on skeleton node data |
| CN103246884A (en) * | 2013-05-22 | 2013-08-14 | 清华大学 | Real-time human body action recognizing method and device based on depth image sequence |
-
2013
- 2013-09-18 CN CN201310428277.XA patent/CN103489000A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20100215257A1 (en) * | 2009-02-25 | 2010-08-26 | Honda Motor Co., Ltd. | Capturing and recognizing hand postures using inner distance shape contexts |
| CN102749993A (en) * | 2012-05-30 | 2012-10-24 | 无锡掌游天下科技有限公司 | Motion recognition method based on skeleton node data |
| CN103246884A (en) * | 2013-05-22 | 2013-08-14 | 清华大学 | Real-time human body action recognizing method and device based on depth image sequence |
Non-Patent Citations (2)
| Title |
|---|
| 洪琛: "交互式系统中三维人体动作识别的研究", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
| 蔡攀: "基于增强现实的实时交互技术研究与实现", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103942526A (en) * | 2014-01-17 | 2014-07-23 | 山东省科学院情报研究所 | Linear feature extraction method for discrete data point set |
| CN103942526B (en) * | 2014-01-17 | 2017-02-08 | 山东省科学院情报研究所 | Linear feature extraction method for discrete data point set |
| CN104317386A (en) * | 2014-06-25 | 2015-01-28 | 西南科技大学 | Action recognition method of posture sequence finite-state machine |
| CN104317386B (en) * | 2014-06-25 | 2017-08-04 | 西南科技大学 | A kind of posture sequence finite state machine action identification method |
| CN104091053A (en) * | 2014-06-26 | 2014-10-08 | 李南君 | Method and equipment for automatically detecting behavior pattern |
| CN104091053B (en) * | 2014-06-26 | 2017-09-29 | 李南君 | Method and apparatus for automatic detection behavior pattern |
| CN106203484A (en) * | 2016-06-29 | 2016-12-07 | 北京工业大学 | A kind of human motion state sorting technique based on classification layering |
| CN106203484B (en) * | 2016-06-29 | 2019-06-21 | 北京工业大学 | A kind of human motion state classification method based on classification layering |
| CN106203503A (en) * | 2016-07-08 | 2016-12-07 | 天津大学 | A kind of action identification method based on skeleton sequence |
| CN106203503B (en) * | 2016-07-08 | 2019-04-05 | 天津大学 | A kind of action identification method based on bone sequence |
| CN106650687A (en) * | 2016-12-30 | 2017-05-10 | 山东大学 | Posture correction method based on depth information and skeleton information |
| CN107225573A (en) * | 2017-07-05 | 2017-10-03 | 上海未来伙伴机器人有限公司 | The method of controlling operation and device of robot |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109558832B (en) | Human body posture detection method, device, equipment and storage medium | |
| CN103489000A (en) | Achieving method of human movement recognition training system | |
| Yu et al. | Mixed pooling for convolutional neural networks | |
| CN104463250B (en) | A kind of Sign Language Recognition interpretation method based on Davinci technology | |
| CN106960206A (en) | Character identifying method and character recognition system | |
| CN106845440A (en) | A kind of augmented reality image processing method and system | |
| CN108596180A (en) | Parameter identification, the training method of parameter identification model and device in image | |
| CN110532925B (en) | Driver fatigue detection method based on space-time graph convolutional network | |
| CN105740808B (en) | Face identification method and device | |
| CN111126280B (en) | Gesture recognition fusion-based aphasia patient auxiliary rehabilitation training system and method | |
| CN106022251A (en) | Abnormal double-person interaction behavior recognition method based on vision co-occurrence matrix sequence | |
| CN105117740A (en) | Font identification method and device | |
| CN109993130A (en) | One kind being based on depth image dynamic sign language semantics recognition system and method | |
| Xing et al. | Oracle bone inscription detection: a survey of oracle bone inscription detection based on deep learning algorithm | |
| CN111722700A (en) | Man-machine interaction method and man-machine interaction equipment | |
| CN114639152A (en) | Multi-modal voice interaction method, device, equipment and medium based on face recognition | |
| CN103544468B (en) | 3D facial expression recognizing method and device | |
| CN113343898B (en) | Mask shielding face recognition method, device and equipment based on knowledge distillation network | |
| CN104616034B (en) | A kind of smog detection method | |
| CN103473562A (en) | Automatic training and identifying system for specific human body action | |
| Ma et al. | Layn: Lightweight multi-scale attention yolov8 network for small object detection | |
| Luo et al. | Multi-scale face detection based on convolutional neural network | |
| CN110956097A (en) | Method and module for extracting occluded human body and method and device for scene conversion | |
| CN115565186B (en) | Training method and device for character recognition model, electronic equipment and storage medium | |
| Fakhfakh et al. | Gesture recognition system for isolated word sign language based on key-point trajectory matrix |
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: 20140101 |
|
| RJ01 | Rejection of invention patent application after publication |