CN108875614A - It is a kind of that detection method is fallen down based on deep learning image procossing - Google Patents
It is a kind of that detection method is fallen down based on deep learning image procossing Download PDFInfo
- Publication number
- CN108875614A CN108875614A CN201810578548.2A CN201810578548A CN108875614A CN 108875614 A CN108875614 A CN 108875614A CN 201810578548 A CN201810578548 A CN 201810578548A CN 108875614 A CN108875614 A CN 108875614A
- Authority
- CN
- China
- Prior art keywords
- human body
- image procossing
- data
- model
- deep learning
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/20—Movements or behaviour, e.g. gesture recognition
- G06V40/23—Recognition of whole body movements, e.g. for sport training
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- General Physics & Mathematics (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Biophysics (AREA)
- Molecular Biology (AREA)
- Biomedical Technology (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Computational Linguistics (AREA)
- Software Systems (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Psychiatry (AREA)
- Social Psychology (AREA)
- Human Computer Interaction (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
Detection method is fallen down based on deep learning image procossing the invention discloses a kind of, this method is not against wearable device and sensor, the picture of the high frequency time shot by camera passes to server end, server end carries out human body critical point detection by Deepcut deep neural network model, the human body critical point detection diagram data of output is input in deep neural network, judgement is fallen down by the model that the training data that key point in all kinds of situations of cut-and-dried human body is distributed trains, every picture processing speed was at 0.2 second or so in terms of image procossing, with very strong real-time.By the above-mentioned means, the present invention can effectively detect fall events.Falling down for different conditions shows that fall events can be effectively detected in the method for proposition with other each form example detections of human body.The present invention can apply in the intelligent domestic system of smart city, ensure the home safety of old man.
Description
Technical field
The present invention relates to health monitoring fields, fall down detection side based on deep learning image procossing more particularly to a kind of
Method.
Background technique
China Today is just being run into aging society, 60 years old or more old man's quantity increasing fast, solitary, lonely
" Empty nest elderly " is also just increasing at an unprecedented rate.It expects 2020,60 years old national or more elderly population will increase
2.55 hundred million people or so are added to, total population specific gravity is accounted for and is promoted to 17.8% or so；Aged will be added to 29,000,000 people
Left and right.As the parent of first generation only child enters old age successively, Chinese tradition family structure is gradually disintegrated, correlative study
It estimates that the year two thousand thirty China's Empty nest elderly will be added to more than 200,000,000, accounts for the ninety percent of the Chinese aged.Old solitary people is on the one hand
Lack spiritual consolation, another aspect ability to act is limited, is even more difficult to notify sons and daughters or parent in time in case of unexpected
People.Old solitary people accidental death in home event takes place frequently in recent years, if correlative study shows that old man can be helped in time after falling down
It helps, 80% mortality risk and 26% long-term treatment risk of being hospitalized can be effectively reduced.
Detection research or fewer is fallen down based on camera under environment of internet of things both at home and abroad, is based on depth
The detection method of falling down for practising image procossing is even more not have.Traditional research for falling down detection, mainstay are all based on
Wearable device and related sensor obtain data by wearable device and are further processed, and use different types of biography
Sense equipment and related algorithm, which are constructed, falls down detection system for different concrete applications.Based on wearable device, environmental sensor
Fall down detection system with high invasive, low precision, poor robustness and it is affected by environment big the disadvantages of.Recently as meter
The detection of falling down of the continuous development of calculation machine vision and digital image processing techniques, view-based access control model becomes a kind of important method.Manually
Fast development and machine learning, the deep learning research of intelligence deepen continuously, and deep learning is with the obvious advantage in terms of intelligence,
It allows and more intelligentized fall down detection and be possibly realized.
Summary of the invention
Detection method is fallen down based on deep learning image procossing the invention mainly solves the technical problem of providing a kind of,
To improve existing the deficiencies of falling down the high and low precision of detection method application cost, poor robustness.
The technical scheme is that：
（1）The candidate region of human part is proposed using deep neural network model DeeperCut, each candidate region is as one
A node, all nodes form an intensive connection figure, and the relevance between node is as the weight between node of graph, by it
As an optimization problem, the component of the same person will be belonged to（Node）It is classified as one kind, everyone is as an independent class.Pass through
DeeperCut model inspection go out 14 key points of human body so that it is determined that people posture.
（2）By the multitude of video prerecorded, including a variety of cameras, multiple angles, several scenes and a variety of
Human body behavior, human body behavior include the falling down and squat down, bend over of various postures, sit the doubtful action behavior fallen down such as lie, will regard
Frequency is cut into ten thousand picture of tale by frame.
（3）These pictures are predicted by DeeperCut model, obtain the number of tens of thousands of label human body key points
According to each of them data include the position coordinates and confidence of 14 points, put on again to each data later and fall down or do not have
There is the label fallen down.
（4）This ten thousand a plurality of data is trained by deep neural network model using keras, output model,
Judgement is fallen down to subsequent picture with model again.Wherein Dense Layer is common full articulamentum, the operation realized
It is output=activation (dot (input, kernel)+bias), wherein activation is swashing by element calculating
Function living, kernel is the weight matrix of this layer, and bias is bias vector.It is tanh in the activation primitive that first five layer uses, and
The last layer has used softmax function to classify.
This method mainly utilizes the thought of Faster RCNN, in characteristic extraction part, using feature relatively advanced at present
Network Resnet152 is extracted, this ensure that feature extraction, that is, the accuracy of detection human body parts first, and wait finding
The feature pyramid network for selecting frame portion point to use also can greatly improve detection speed.
It is provided in an embodiment of the present invention that detection method is fallen down based on deep learning image procossing, it is shot by camera
The picture of high frequency time passes to server end, and server end carries out the inspection of human body key point by Deepcut deep neural network model
It surveys, the human body critical point detection diagram data of output is input in deep neural network, all kinds of feelings of cut-and-dried human body are passed through
The model that the training data that key point is distributed under condition trains falls down judgement, every picture processing speed in terms of image procossing
At 0.2 second or so, there is very strong real-time.
The beneficial effects of the invention are as follows：Rapidly and accurately detect fall events.
Detailed description of the invention
Fig. 1 is the system framework schematic diagram of case study on implementation of the present invention；
Fig. 2 shows a kind of distribution schematic diagrams of human body key point；
Fig. 3 shows the actual human body critical point detection effect in embodiment；
Fig. 4 shows the deep neural network structure in embodiment；
Fig. 5 shows training accuracy rate variation diagram of the training pattern in test data in embodiment；
Fig. 6 shows the result that part in embodiment falls down judgement to single picture.
Specific embodiment
In order to make the objectives, technical solutions, and advantages of the present invention clearer, right below in conjunction with drawings and examples
The present invention is further elaborated, it should be understood that and the specific embodiments described herein are merely illustrative of the present invention, and
It is not used in the restriction present invention
Shown in referring to Fig.1, the present invention is not against wearable device and sensor, the picture service of passing to of the high frequency time of camera shooting
Device end, server end are fallen down judgement, are fallen down if detecting, are given notice by wechat or short message to household.Specifically include with
Lower step：
S1：By DeeperCut model inspection go out human body 14 key points so that it is determined that people posture.
S2：Multitude of video is prerecorded, including a variety of cameras, multiple angles, several scenes, a variety of human body rows
Include the falling down and squat down, bend over of various postures for, human body behavior, sit the doubtful action behavior fallen down such as lie, video is pressed into frame
It is cut into ten thousand picture of tale.
S3：These pictures are predicted by DeeperCut model, obtain the number of tens of thousands of label human body key points
According to each of them data include the position coordinates and confidence of 14 points, put on again to each data later and fall down or do not have
There is the label fallen down.
S4：This ten thousand a plurality of data is trained by deep neural network model using keras, output model.
S5：Judgement is fallen down to subsequent picture with model.
As described in step S1,14 key points of human body are distributed as shown in Fig. 2, actually detected effect is as shown in figure 3,14
Key point can determine the posture of people.
As described in step S2, the generalization ability of training pattern can be improved in a large amount of picture, so that more accurately judgement is fallen
?.
As described in step S3, the coordinate data and confidence of a series of a large amount of available human body key points of picture will be every
Picture marks, and trains so as to input neural network.
As described in step S4, Fig. 4 shows the structure of deep neural network, and wherein Dense Layer is common complete
Articulamentum, the operation realized is output=activation (dot (input, kernel)+bias), wherein
Activation is the activation primitive calculated by element, and kernel is the weight matrix of this layer, and bias is bias vector.At first five
The activation primitive that layer uses is tanh, and the last layer has used softmax function to classify.Training data is inputted into mind
Through network, output model, Fig. 5 shows training accuracy rate variation diagram of the training pattern in test data, and final accuracy rate is steady
It is scheduled on 98% or so.
As described in step S5, judgement fallen down to subsequent pictures using model, Fig. 6 shows partial detection, in Fig. 6
Fall is to detect tumble, and tumble is not detected in NOT FALL expression.
Claims (3)
1. a kind of fall down detection method based on deep learning image procossing, it is characterised in that：Not against wearable device and sensing
Device is that a kind of picture of high frequency time falling down detection method based on deep learning image procossing, being shot using camera is passed to
Server end, server end carry out human body critical point detection by Deepcut deep neural network model, and the human body of output is closed
Key point detection diagram data is input in deep neural network, the instruction being distributed by key point in all kinds of situations of cut-and-dried human body
The model that white silk data train is to judge whether to fall down.
A kind of detection method is fallen down based on deep learning image procossing 2. according to claim 1, it is characterised in that：With
Tens of thousands of pictures go out 14 key points of human body in figure by DeeperCut model prediction as training data, obtain tens of thousands of
The data of human body key point are marked, each of them data include the position coordinates and confidence of 14 points, give each again later
Data put on the label fallen down again without falling down.
A kind of detection method is fallen down based on deep learning image procossing 3. according to claim 2, it is characterised in that：Make
This ten thousand a plurality of data is trained by deep neural network model with keras, output model, then with model to subsequent
Picture fall down judgement, wherein Dense Layer is common full articulamentum, the operation realized be output=
Activation (dot (input, kernel)+bias), wherein activation is the activation primitive calculated by element,
Kernel is the weight matrix of this layer, and bias is bias vector, is tanh in the activation primitive that first five layer uses, and the last layer
Softmax function has been used to classify.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810578548.2A CN108875614A (en) | 2018-06-07 | 2018-06-07 | It is a kind of that detection method is fallen down based on deep learning image procossing |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810578548.2A CN108875614A (en) | 2018-06-07 | 2018-06-07 | It is a kind of that detection method is fallen down based on deep learning image procossing |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108875614A true CN108875614A (en) | 2018-11-23 |
Family
ID=64337224
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810578548.2A Pending CN108875614A (en) | 2018-06-07 | 2018-06-07 | It is a kind of that detection method is fallen down based on deep learning image procossing |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108875614A (en) |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109740446A (en) * | 2018-12-14 | 2019-05-10 | 深圳壹账通智能科技有限公司 | Classroom students ' behavior analysis method and device |
| CN109919077A (en) * | 2019-03-04 | 2019-06-21 | 网易（杭州）网络有限公司 | Gesture recognition method, device, medium and calculating equipment |
| CN111209848A (en) * | 2020-01-03 | 2020-05-29 | 北京工业大学 | Real-time fall detection method based on deep learning |
| CN111563397A (en) * | 2019-02-13 | 2020-08-21 | 阿里巴巴集团控股有限公司 | Detection method, detection device, intelligent equipment and computer storage medium |
| CN113221661A (en) * | 2021-04-14 | 2021-08-06 | 浪潮天元通信信息系统有限公司 | Intelligent human body tumbling detection system and method |
| CN113936425A (en) * | 2021-09-26 | 2022-01-14 | 上海深豹智能科技有限公司 | Home-based old-age falling detection system and operation method |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104850846A (en) * | 2015-06-02 | 2015-08-19 | 深圳大学 | Human behavior recognition method and human behavior recognition system based on depth neural network |
| CN107220604A (en) * | 2017-05-18 | 2017-09-29 | 清华大学深圳研究生院 | A kind of fall detection method based on video |
-
2018
- 2018-06-07 CN CN201810578548.2A patent/CN108875614A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104850846A (en) * | 2015-06-02 | 2015-08-19 | 深圳大学 | Human behavior recognition method and human behavior recognition system based on depth neural network |
| CN107220604A (en) * | 2017-05-18 | 2017-09-29 | 清华大学深圳研究生院 | A kind of fall detection method based on video |
Non-Patent Citations (1)
| Title |
|---|
| ELDAR INSAFUTDINOV等: "DeeperCut：A Deeper，Stronger，and Faster Multi-Person Pose Estimation Model", 《ARXIV》 * |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109740446A (en) * | 2018-12-14 | 2019-05-10 | 深圳壹账通智能科技有限公司 | Classroom students ' behavior analysis method and device |
| CN111563397A (en) * | 2019-02-13 | 2020-08-21 | 阿里巴巴集团控股有限公司 | Detection method, detection device, intelligent equipment and computer storage medium |
| CN109919077A (en) * | 2019-03-04 | 2019-06-21 | 网易（杭州）网络有限公司 | Gesture recognition method, device, medium and calculating equipment |
| CN111209848A (en) * | 2020-01-03 | 2020-05-29 | 北京工业大学 | Real-time fall detection method based on deep learning |
| CN113221661A (en) * | 2021-04-14 | 2021-08-06 | 浪潮天元通信信息系统有限公司 | Intelligent human body tumbling detection system and method |
| CN113936425A (en) * | 2021-09-26 | 2022-01-14 | 上海深豹智能科技有限公司 | Home-based old-age falling detection system and operation method |
| CN113936425B (en) * | 2021-09-26 | 2024-03-26 | 上海深豹智能科技有限公司 | Fall detection system for home-based aged people and operation method |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108875614A (en) | It is a kind of that detection method is fallen down based on deep learning image procossing | |
| Tsai et al. | Implementation of fall detection system based on 3D skeleton for deep learning technique | |
| CN104992167B (en) | A kind of method for detecting human face and device based on convolutional neural networks | |
| CN109101865A (en) | A kind of recognition methods again of the pedestrian based on deep learning | |
| CN108764085A (en) | Based on the people counting method for generating confrontation network | |
| CN113111767A (en) | Fall detection method based on deep learning 3D posture assessment | |
| CN111488850B (en) | Neural network-based old people falling detection method | |
| Yu et al. | Railway obstacle detection algorithm using neural network | |
| CN110232379A (en) | A kind of vehicle attitude detection method and system | |
| CN110378179A (en) | Subway based on infrared thermal imaging is stolen a ride behavioral value method and system | |
| CN109918995B (en) | Crowd abnormity detection method based on deep learning | |
| CN113642361A (en) | Method and equipment for detecting falling behavior | |
| CN106780565A (en) | A kind of many students based on light stream and k means clusters rise and sit detection method | |
| CN110751063A (en) | Infant quilt kicking prevention recognition device and method based on deep learning | |
| CN109522961A (en) | A kind of semi-supervision image classification method based on dictionary deep learning | |
| Zhang et al. | Fall detection in videos with trajectory-weighted deep-convolutional rank-pooling descriptor | |
| Shen et al. | Fall detection system based on deep learning and image processing in cloud environment | |
| CN110096945A (en) | Indoor Video key frame of video real time extracting method based on machine learning | |
| Tang et al. | Intelligent video surveillance system for elderly people living alone based on ODVS | |
| Yao et al. | An improved feature-based method for fall detection | |
| CN107256566B (en) | Forest fires detection method based on radiation energy | |
| Sun et al. | Sitting posture recognition in real-time combined with index map and BLS | |
| Dong et al. | CCTwins: A Weakly Supervised Transformer-Based Crowd Counting Method With Adaptive Scene Consistency Attention | |
| CN114913585A (en) | Household old man falling detection method integrating facial expressions | |
| Lin et al. | People recognition in multi-cameras using the visual color processing mechanism |
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 | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20181123 |