CN108446594A - Emergency reaction ability assessment method based on action recognition - Google Patents
Emergency reaction ability assessment method based on action recognition Download PDFInfo
- Publication number
- CN108446594A CN108446594A CN201810142115.2A CN201810142115A CN108446594A CN 108446594 A CN108446594 A CN 108446594A CN 201810142115 A CN201810142115 A CN 201810142115A CN 108446594 A CN108446594 A CN 108446594A
- Authority
- CN
- China
- Prior art keywords
- personnel
- sequence
- emergency
- demonstration
- neural networks
- 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.)
- Granted
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
-
- 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
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
Abstract
The invention discloses a kind of emergency reaction ability assessment method based on action recognition, includes the following steps:Demonstration personnel carry out emergency action demonstration, acquire motion images sequence;Motion images sequence is converted into the motion characteristic sequence vector of CNN neural networks output;Motion characteristic sequence vector is exported to the general characteristic vector of emergency action sequence by RNN neural networks;Tested personnel shows the emergency action consistent with demonstration personnel, and same way is used in combination to obtain the general characteristic vector of emergency action sequence;The difference between demonstration personnel and the general characteristic vector of tested personnel is calculated, the emergency reaction ability that tested personnel is carried out according to difference size is evaluated.The present invention is by CNN neural networks and RNN Application of Neural Network in emergency reaction ability assessment process, the emergency capability test and appraisal of tested personnel can be rapidly completed in batches, and it is evaluating result objective and fair, scientific and effective, it is particularly suitable for the emergency reaction ability test and appraisal when facing natural calamity, artificial injury.
Description
Technical field
The present invention relates to a kind of emergency reaction ability assessment methods, more particularly to one kind facing natural calamity, artificial wound
The emergency reaction ability assessment method based on action recognition when evil.
Background technology
When facing natural calamity and artificial injure, only appropriate quickly emergency reaction can just evade danger, protection certainly
Oneself.And in safety education and training, how scientific and effective, quantization effectively to be carried out intuitively to the emergency reaction ability of tested personnel
Test and appraisal, just become one of the key link of training.
The current safety education and training level of informatization is relatively low, it is also necessary to which teacher of the cultivation is by directly observing the dynamic of tested personnel
Make reaction to make the judgement of output capacity relative superiority or inferiority, therefore less efficient, and subjectivity is strong, lacks science accurately assessment approach.
Application No. is the applications for a patent for invention of " 201310164059.X ", disclose a kind of be suitable under fuzzy uncertain environment
Truck driver's emergency reaction ability evaluation method, step include:Determine evaluation group and decision index system value;Determine that index is weighed
Weight;Calculate normalization decision matrix;It determines and obscures positive ideal solution and fuzzy minus ideal result;Calculate truck driver's emergency reaction ability
The distance between evaluation of estimate and positive ideal solution and minus ideal result;The relative similarity degree and sequence for calculating and obscuring positive ideal solution.
The specific method of foregoing invention application, and evaluating result is obtained by the artificial test and appraisal of panel of expert, equally deposit
Subjectivity it is strong, lack the problem of scientific basis.
Current depth learning art has been achieved for a series of breakthrough progress in image processing field, including convolution is refreshing
Through network C NN, Recognition with Recurrent Neural Network RNN, the model of many standards in the industry is all formd, as the fields CNN are typical
The typical LSTM networks in the field GoogLeNet, RNN, these network structures relative maturity and have standardized, and can handle various figures
As problem, but the test and appraisal field based on action it is not applied at present.
Invention content
The purpose of the present invention is that solve the above-mentioned problems and provides one kind and being applied to CNN neural networks and RNN god
The emergency reaction ability assessment method based on action recognition through network.
The present invention is achieved through the following technical solutions above-mentioned purpose:
A kind of emergency reaction ability assessment method based on action recognition, includes the following steps:
Step 1:Demonstration personnel carry out emergency action demonstration, and the motion images sequence of demonstration personnel is acquired using camera;
Step 2:The motion images sequence for the personnel of demonstrating is sequentially input into CNN neural networks, by the action diagram for the personnel of demonstrating
As sequence is converted into the motion characteristic sequence vector of the demonstration personnel of CNN neural networks output;CNN neural networks, that is, convolutional Neural
Network is a kind of feedforward neural network, and artificial neuron can respond surrounding cells, can large-scale image procossing;
Step 3:The motion characteristic sequence vector for the personnel of demonstrating is sequentially input RNN neural networks to be trained, RNN nerves
Network output is next motion characteristic vector of prediction demonstration personnel, the motion characteristic vector input of the last one demonstration personnel
After RNN neural networks, corresponding output is the general characteristic vector of the emergency action sequence of demonstration personnel;RNN neural networks
That is Recognition with Recurrent Neural Network, the speech recognition being widely used in natural language processing, hand-written book is not and the fields such as machine translation;
Step 4:Tested personnel shows the emergency action consistent with demonstration personnel, and the dynamic of tested personnel is acquired using camera
Make image sequence;
Step 5:The motion images sequence of tested personnel is sequentially input into CNN neural networks, by the action diagram of tested personnel
As sequence is converted into the motion characteristic sequence vector of the tested personnel of CNN neural networks output;
Step 6:The motion characteristic sequence vector of tested personnel is sequentially input RNN neural networks to be trained, RNN nerves
Network output is next motion characteristic vector of prediction tested personnel, the motion characteristic vector input of the last one tested personnel
After RNN neural networks, corresponding output is the general characteristic vector of the emergency action sequence of tested personnel;
Step 7:Calculate the difference between the general characteristic vector of the general characteristic vector sum tested personnel of demonstration personnel, root
The emergency reaction ability evaluation of tested personnel is carried out according to difference size.
Preferably, in the step 1, the frame per second of the camera is 24 frames/second, resolution 1024*768, continuously
Acquisition 10 minutes, i.e. every group of motion images sequence include 14400 pictures, and times of collection is 1000 times, i.e. 1000 groups of action diagrams
As sequence.
Preferably, in the step 2 and the step 5, the CNN neural networks are GoogLeNet networks;
Googlenet is a kind of depth convolutional neural networks model of google designs, and first edition depth can be to 22 layers, this network is adopted
It has received the thought of sparse study, network size is increased by the parameter of sparse network.Further, the step 2 and the step
In rapid 5, the CNN neural networks are the standard Inception-V4 networks in GoogLeNet networks.GoogLeNet is to network
In traditional convolutional layer be modified, it is proposed that be referred to as the structure of Inception, for increasing network depth and width,
Improve deep neural network performance;InceptionV4 compared to InceptionV3 mainly in conjunction with the ResNet of Microsoft, will be wrong
Accidentally rate is further reduced to 3.08%.
Preferably, in the step 3, the RNN neural networks are the single layer LSTM networks of standard, and Hidden nodes are
128, using ADAM optimization algorithms, momentum 0.5, initial learning rate 0.0002, per 100 decaying half of iteration, every time training
64 motion images sequences are inputted simultaneously as a batch.LSTM (Long Short-Term Memory) is shot and long term memory
Network is a kind of Recognition with Recurrent Neural Network, is suitable for being spaced and postpone relatively long important thing in processing and predicted time sequence
Part;Adam optimization algorithms are a kind of to optimize the algorithm of random targets function based on First-order Gradient.
Preferably, in the step 4, the frame per second of the camera is 24 frames/second, resolution 1024*768, continuously
Acquisition 10 minutes, i.e. every group of motion images sequence include 14400 pictures, and times of collection is 1 time, i.e. 1 group of motion images sequence.
Preferably, in the step 6, the RNN neural networks are the motion images sequence training by demonstration number
Good LSTM networks, network parameter are constant.
Preferably, in the step 7, the calculation formula of the general characteristic vector difference is similar using the cosine of standard
Spend calculation formula.
The beneficial effects of the present invention are:
The present invention by CNN neural networks and RNN Application of Neural Network in emergency reaction ability assessment process, can batch
The emergency capability test and appraisal of tested personnel are rapidly completed, and evaluating result objective and fair, scientific and effective, are particularly suitable for facing
Emergency reaction ability test and appraisal when natural calamity, artificial injury.
Specific implementation mode
With reference to embodiment, the invention will be further described:
Embodiment:
Below by taking seismic emergency responding capability comparison as an example:
A kind of emergency reaction ability assessment method based on action recognition, includes the following steps:
Step 1:Camera is disposed in earthquake emergency test site first, is existed to acquire demonstration personnel and tested personnel
When simulating earthquake generation, the emergency reaction action sequence within 10 minutes;
Step 2:Demonstration personnel carry out emergency action demonstration, such as quickly hide the desk to test site in the following, both hands are handed over
Fork is put into protect head etc. after neck, and the motion images sequence of demonstration personnel is acquired using camera;Wherein, the frame of camera
Rate is 24 frames/second, and resolution 1024*768, continuous acquisition 10 minutes, i.e. every group of motion images sequence include 14400 figures
Piece, times of collection are 1000 times, i.e. 1000 groups of motion images sequences;
Step 3:The motion images sequence for the personnel of demonstrating is sequentially input in the GoogLeNet networks in CNN neural networks
Standard Inception-V4 networks, by the motion images sequence for the personnel of demonstrating be converted into Inception-V4 networks output show
The motion characteristic sequence vector of model personnel;In application, remove softmax0, softmax1 in standard GoogLeNet networks,
Softmax2, softmaxActivation;
Step 4:The motion characteristic sequence vector for the personnel of demonstrating is sequentially input to the single layer LSTM networks in RNN neural networks
It is trained, Hidden nodes are 128, using ADAM optimization algorithms, momentum 0.5, initial learning rate 0.0002, per iteration 100
Secondary decaying half trains while inputting 64 motion images sequences as a batch, trained all samples primary every time
Iteration, altogether repetitive exercise 600 times;The output of single layer LSTM networks is vectorial for next motion characteristic of prediction demonstration personnel, most
After the motion characteristic vector input single layer LSTM networks of the latter demonstration personnel, corresponding output as demonstration personnel's answers jerking movement
Make the general characteristic vector of sequence;
Step 5:Tested personnel shows the emergency action consistent with demonstration personnel, and the dynamic of tested personnel is acquired using camera
Make image sequence;Wherein, the frame per second of camera is 24 frames/second, resolution 1024*768, continuous acquisition 10 minutes, i.e., every group
Motion images sequence includes 14400 pictures, and times of collection is 1 time, i.e. 1 group of motion images sequence;
Step 6:The motion images sequence of tested personnel is sequentially input in the GoogLeNet networks in CNN neural networks
Standard Inception-V4 networks, by the motion images sequence of tested personnel be converted into Inception-V4 networks output quilt
The motion characteristic sequence vector of survey personnel;
Step 7:The motion characteristic sequence vector of tested personnel is sequentially input into the motion images sequence by number of demonstrating
Trained LSTM networks are trained, and network parameter is constant, and LSTM networks output is the next dynamic of prediction tested personnel
Make feature vector, after the motion characteristic vector of the last one tested personnel inputs the LSTM networks, corresponding output is tested
The general characteristic vector of the emergency action sequence of personnel;
Step 8:Calculate the difference between the general characteristic vector of the general characteristic vector sum tested personnel of demonstration personnel, root
The emergency reaction ability evaluation of tested personnel is carried out according to difference size;The calculation formula of general characteristic vector difference is using standard
Cosine similarity calculation formula, it is specific as follows:
If the general characteristic vector of the emergency action sequence of the personnel of demonstration is indicated with a, the emergency action sequence of tested personnel
General characteristic vector indicated with b, | | a | | indicate the mould of vector a, | | b | | the mould of expression vector b, then cosine similarity cos θ
Calculation formula is as follows:
The magnitude range of cosine similarity is [- 1,1], and value is bigger, illustrates that the action sequence of tested personnel is closer and shows
Model personnel, therefore its seismic emergency responding ability is stronger.
Above-described embodiment is presently preferred embodiments of the present invention, is not the limitation to technical solution of the present invention, as long as
Without the technical solution that creative work can be realized on the basis of the above embodiments, it is regarded as falling into patent of the present invention
Rights protection scope in.
Claims (8)
1. a kind of emergency reaction ability assessment method based on action recognition, it is characterised in that:Include the following steps:
Step 1:Demonstration personnel carry out emergency action demonstration, and the motion images sequence of demonstration personnel is acquired using camera;
Step 2:The motion images sequence for the personnel of demonstrating is sequentially input into CNN neural networks, by the motion images sequence for the personnel of demonstrating
Row are converted into the motion characteristic sequence vector of the demonstration personnel of CNN neural networks output;
Step 3:The motion characteristic sequence vector for the personnel of demonstrating is sequentially input RNN neural networks to be trained, RNN neural networks
Output is vectorial for next motion characteristic of prediction demonstration personnel, the motion characteristic vector of the last one demonstration personnel inputs RNN
After neural network, corresponding output is the general characteristic vector of the emergency action sequence of demonstration personnel;
Step 4:Tested personnel shows the emergency action consistent with demonstration personnel, and the action diagram of tested personnel is acquired using camera
As sequence;
Step 5:The motion images sequence of tested personnel is sequentially input into CNN neural networks, by the motion images sequence of tested personnel
Row are converted into the motion characteristic sequence vector of the tested personnel of CNN neural networks output;
Step 6:The motion characteristic sequence vector of tested personnel is sequentially input RNN neural networks to be trained, RNN neural networks
Output is next motion characteristic vector of prediction tested personnel, the motion characteristic vector input RNN of the last one tested personnel
After neural network, corresponding output is the general characteristic vector of the emergency action sequence of tested personnel;
Step 7:The difference between the general characteristic vector of the general characteristic vector sum tested personnel of demonstration personnel is calculated, according to difference
Different size carries out the emergency reaction ability evaluation of tested personnel.
2. the emergency reaction ability assessment method according to claim 1 based on action recognition, it is characterised in that:The step
In rapid 1, the frame per second of the camera is 24 frames/second, resolution 1024*768, continuous acquisition 10 minutes, i.e. every group of action diagram
Picture sequence includes 14400 pictures, and times of collection is 1000 times, i.e. 1000 groups of motion images sequences.
3. the emergency reaction ability assessment method according to claim 1 based on action recognition, it is characterised in that:The step
Rapid 2 and the step 5 in, the CNN neural networks be GoogLeNet networks.
4. the emergency reaction ability assessment method according to claim 3 based on action recognition, it is characterised in that:The step
Rapid 2 and the step 5 in, the CNN neural networks be GoogLeNet networks in standard Inception-V4 networks.
5. the emergency reaction ability assessment method according to claim 1 based on action recognition, it is characterised in that:The step
In rapid 3, the RNN neural networks are the single layer LSTM networks of standard, and Hidden nodes are 128, using ADAM optimization algorithms, are moved
Amount 0.5, initial learning rate 0.0002, per 100 decaying half of iteration, 64 motion images sequences of training while input are made every time
For a batch.
6. the emergency reaction ability assessment method according to claim 1 based on action recognition, it is characterised in that:The step
In rapid 4, the frame per second of the camera is 24 frames/second, resolution 1024*768, continuous acquisition 10 minutes, i.e. every group of action diagram
Picture sequence includes 14400 pictures, and times of collection is 1 time, i.e. 1 group of motion images sequence.
7. the emergency reaction ability assessment method according to claim 1 based on action recognition, it is characterised in that:The step
In rapid 6, the RNN neural networks are the trained LSTM networks of motion images sequence by demonstration number, and network parameter is not
Become.
8. the emergency reaction ability assessment method according to claim 1 based on action recognition, it is characterised in that:The step
In rapid 7, the calculation formula of the general characteristic vector difference uses the cosine similarity calculation formula of standard.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810142115.2A CN108446594B (en) | 2018-02-11 | 2018-02-11 | Emergency response capability evaluation method based on action recognition |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810142115.2A CN108446594B (en) | 2018-02-11 | 2018-02-11 | Emergency response capability evaluation method based on action recognition |
Publications (2)
Publication Number | Publication Date |
---|---|
CN108446594A true CN108446594A (en) | 2018-08-24 |
CN108446594B CN108446594B (en) | 2021-08-06 |
Family
ID=63192459
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810142115.2A Active CN108446594B (en) | 2018-02-11 | 2018-02-11 | Emergency response capability evaluation method based on action recognition |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108446594B (en) |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110222556A (en) * | 2019-04-22 | 2019-09-10 | 北方工业大学 | A kind of human action identifying system and method |
CN111797778A (en) * | 2020-07-08 | 2020-10-20 | 龙岩学院 | Automatic scoring method for breaking street dance anchor and wheat dance |
US20210110203A1 (en) * | 2019-10-11 | 2021-04-15 | Perceptive Automata, Inc. | Visualizing machine learning predictions of human interaction with vehicles |
CN114783046A (en) * | 2022-03-01 | 2022-07-22 | 北京赛思信安技术股份有限公司 | CNN and LSTM-based human body continuous motion similarity scoring method |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105975956A (en) * | 2016-05-30 | 2016-09-28 | 重庆大学 | Infrared-panorama-pick-up-head-based abnormal behavior identification method of elderly people living alone |
US20170255832A1 (en) * | 2016-03-02 | 2017-09-07 | Mitsubishi Electric Research Laboratories, Inc. | Method and System for Detecting Actions in Videos |
CN107220611A (en) * | 2017-05-23 | 2017-09-29 | 上海交通大学 | A kind of space-time feature extracting method based on deep neural network |
-
2018
- 2018-02-11 CN CN201810142115.2A patent/CN108446594B/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20170255832A1 (en) * | 2016-03-02 | 2017-09-07 | Mitsubishi Electric Research Laboratories, Inc. | Method and System for Detecting Actions in Videos |
CN105975956A (en) * | 2016-05-30 | 2016-09-28 | 重庆大学 | Infrared-panorama-pick-up-head-based abnormal behavior identification method of elderly people living alone |
CN107220611A (en) * | 2017-05-23 | 2017-09-29 | 上海交通大学 | A kind of space-time feature extracting method based on deep neural network |
Non-Patent Citations (2)
Title |
---|
ZHONGXU HU 等: "Hand pose estimation with CNN-RNN", 《2017 EUROPEAN CONFERENCE ON ELECTRICAL ENGINEERING AND COMPUTER SCIENCE》 * |
朱煜 等: "基于深度学习的人体行为识别算法综述", 《自动化学报》 * |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110222556A (en) * | 2019-04-22 | 2019-09-10 | 北方工业大学 | A kind of human action identifying system and method |
US20210110203A1 (en) * | 2019-10-11 | 2021-04-15 | Perceptive Automata, Inc. | Visualizing machine learning predictions of human interaction with vehicles |
US11551030B2 (en) * | 2019-10-11 | 2023-01-10 | Perceptive Automata, Inc. | Visualizing machine learning predictions of human interaction with vehicles |
CN111797778A (en) * | 2020-07-08 | 2020-10-20 | 龙岩学院 | Automatic scoring method for breaking street dance anchor and wheat dance |
CN111797778B (en) * | 2020-07-08 | 2023-06-02 | 龙岩学院 | Automatic scoring method for break-in street dance and wheat-linking dancing |
CN114783046A (en) * | 2022-03-01 | 2022-07-22 | 北京赛思信安技术股份有限公司 | CNN and LSTM-based human body continuous motion similarity scoring method |
Also Published As
Publication number | Publication date |
---|---|
CN108446594B (en) | 2021-08-06 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108615048B (en) | Defense method for image classifier adversity attack based on disturbance evolution | |
CN108446594A (en) | Emergency reaction ability assessment method based on action recognition | |
CN108229582A (en) | Entity recognition dual training method is named in a kind of multitask towards medical domain | |
CN104598611B (en) | The method and system being ranked up to search entry | |
CN111159419B (en) | Knowledge tracking data processing method, system and storage medium based on graph convolution | |
CN112818764B (en) | Low-resolution image facial expression recognition method based on feature reconstruction model | |
CN111860267B (en) | Multichannel body-building exercise identification method based on human body skeleton joint point positions | |
CN106886572A (en) | Knowledge mapping relationship type estimation method and its device based on Markov Logic Networks | |
CN110807509A (en) | Depth knowledge tracking method based on Bayesian neural network | |
CN113361685B (en) | Knowledge tracking method and system based on learner knowledge state evolution expression | |
CN110826056A (en) | Recommendation system attack detection method based on attention convolution self-encoder | |
CN115545160B (en) | Knowledge tracking method and system for multi-learning behavior collaboration | |
Termritthikun et al. | Accuracy improvement of Thai food image recognition using deep convolutional neural networks | |
CN110826459B (en) | Migratable campus violent behavior video identification method based on attitude estimation | |
CN108038467A (en) | The sparse face identification method that a kind of mirror image is combined with thickness level | |
CN112884150A (en) | Safety enhancement method for knowledge distillation of pre-training model | |
Liang et al. | Out-of-distribution generalization with deep equilibrium models | |
Fidanova et al. | InterCriteria analysis of different metaheuristics applied to E. coli cultivation process | |
CN113378985A (en) | Countermeasure sample detection method and device based on layer-by-layer correlation propagation | |
Syaharuddin et al. | Accuracy rate of ANN back propagation architecture with modified algorithm: A meta-analysis | |
Zhao et al. | Research on Deep Knowledge Tracing Model Integrating Graph Attention Network | |
CN117238026B (en) | Gesture reconstruction interactive behavior understanding method based on skeleton and image features | |
Dang | Development of Structural Damage Detection Method Working with Contaminated Vibration Data via Autoencoder and Gradient Boosting | |
CN113762082B (en) | Unsupervised skeleton action recognition method based on cyclic graph convolution automatic encoder | |
Wang | Evaluation Method of the Influence of Sports Training on Physical Index Based on Deep Learning |
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 |