CN106650644B - The recognition methods of driver's hazardous act and system - Google Patents
The recognition methods of driver's hazardous act and system Download PDFInfo
- Publication number
- CN106650644B CN106650644B CN201611116789.2A CN201611116789A CN106650644B CN 106650644 B CN106650644 B CN 106650644B CN 201611116789 A CN201611116789 A CN 201611116789A CN 106650644 B CN106650644 B CN 106650644B
- Authority
- CN
- China
- Prior art keywords
- hazardous act
- forest
- hazardous
- matrix
- driver
- 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
Links
Classifications
-
- 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
-
- B—PERFORMING OPERATIONS; TRANSPORTING
- B60—VEHICLES IN GENERAL
- B60W—CONJOINT CONTROL OF VEHICLE SUB-UNITS OF DIFFERENT TYPE OR DIFFERENT FUNCTION; CONTROL SYSTEMS SPECIALLY ADAPTED FOR HYBRID VEHICLES; ROAD VEHICLE DRIVE CONTROL SYSTEMS FOR PURPOSES NOT RELATED TO THE CONTROL OF A PARTICULAR SUB-UNIT
- B60W40/00—Estimation or calculation of non-directly measurable driving parameters for road vehicle drive control systems not related to the control of a particular sub unit, e.g. by using mathematical models
- B60W40/08—Estimation or calculation of non-directly measurable driving parameters for road vehicle drive control systems not related to the control of a particular sub unit, e.g. by using mathematical models related to drivers or passengers
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
- G01S15/00—Systems using the reflection or reradiation of acoustic waves, e.g. sonar systems
- G01S15/88—Sonar systems specially adapted for specific applications
-
- B—PERFORMING OPERATIONS; TRANSPORTING
- B60—VEHICLES IN GENERAL
- B60W—CONJOINT CONTROL OF VEHICLE SUB-UNITS OF DIFFERENT TYPE OR DIFFERENT FUNCTION; CONTROL SYSTEMS SPECIALLY ADAPTED FOR HYBRID VEHICLES; ROAD VEHICLE DRIVE CONTROL SYSTEMS FOR PURPOSES NOT RELATED TO THE CONTROL OF A PARTICULAR SUB-UNIT
- B60W40/00—Estimation or calculation of non-directly measurable driving parameters for road vehicle drive control systems not related to the control of a particular sub unit, e.g. by using mathematical models
- B60W40/08—Estimation or calculation of non-directly measurable driving parameters for road vehicle drive control systems not related to the control of a particular sub unit, e.g. by using mathematical models related to drivers or passengers
- B60W2040/0818—Inactivity or incapacity of driver
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/50—Context or environment of the image
- G06V20/59—Context or environment of the image inside of a vehicle, e.g. relating to seat occupancy, driver state or inner lighting conditions
- G06V20/597—Recognising the driver's state or behaviour, e.g. attention or drowsiness
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Radar, Positioning & Navigation (AREA)
- Remote Sensing (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Physics & Mathematics (AREA)
- Transportation (AREA)
- Acoustics & Sound (AREA)
- Computer Networks & Wireless Communication (AREA)
- Mechanical Engineering (AREA)
- Mathematical Physics (AREA)
- Artificial Intelligence (AREA)
- Signal Processing (AREA)
- General Engineering & Computer Science (AREA)
- Automation & Control Theory (AREA)
- Traffic Control Systems (AREA)
Abstract
A kind of driver's hazardous act recognition methods and system, ultrasound field is set in vehicle first, obtain different mode of the hazardous act on Doppler frequency spectrum, then principal character is extracted through principal component analysis, the multi-categorizer of different hazardous acts is identified by algorithm of support vector machine, and gradient former forest is established according to the completeness of hazardous act and duration, real-time ultrasound ripple signal is finally gone out into hazardous act by gradient former Forest mapping after window cutting algorithm is cut into slices and given a warning, the present invention can while accuracy of identification is ensured can hazardous act early stage can recognition result go out and provide alarm, need not rely on any additional external equipment, it is not susceptible to the interference of the external environments such as weather.
Description
Technical field
The present invention relates to a kind of technology in intelligent behaviour identification field, specifically a kind of driver's hazardous act identification
Method and system.
Background technology
Detect the danger of in-car driver or the method for omission is broadly divided into two classes, one kind is by camera, separately
One kind passes through sensor.By the camera on intelligent terminal, image information is extracted to identify omission.By being arranged at car
Sensor on, i.e. accelerometer and gyroscope, sensing data is obtained to identify carelessness or hazardous act.Sensor can only
The state of vehicle is monitored, the behavior state of driver can not be directly obtained, and the requirement to intelligent terminal is too high.
The content of the invention
The present invention is higher for requirement of the prior art to IMAQ in detection process, the defects of robust sex-intergrade,
A kind of driver's hazardous act recognition methods and system are proposed, can be in hazardous act while accuracy of identification is ensured
Early stage can recognition result go out and provide alarm, it is not necessary to rely on any additional external equipment, it is not easy to by weather etc.
The interference of external environment.
The present invention is achieved by the following technical solutions：
The present invention relates to a kind of driver's hazardous act recognition methods, ultrasound field is set in vehicle first, endangered
Different mode of the dangerous behavior on Doppler frequency spectrum, principal character then is extracted through principal component analysis, is calculated by SVMs
Method is identified the multi-categorizer of different hazardous acts, and establishes gradient former according to the completeness and duration of hazardous act
Forest, real-time ultrasound ripple signal is finally gone out into dangerous row after window cutting algorithm is cut into slices by gradient former Forest mapping
For and give a warning.
Described driver's hazardous act recognition methods specifically includes following steps：
1) receive in-car ultrasonic signal and handle, obtain the data matrix X' of ultrasonic signal spectrum structure；
2) d principal character δ={ δ is extracted by principal component analysis to some data matrix X'1,δ2,....δd}；
3) based on principal character δ and supporting vector machine model structure multi-categorizer model θ；
4) the completeness α and duration τ of hazardous act are directed to, gradient former forest is established with reference to multi-categorizer model θ
Θ；
5) by window cutting algorithm by the action data X in duration ττGradient former forest Θ is sent into, is identified
Hazardous act then sends alarm.
Described step 1) specifically includes following steps：
1.1) single-frequency ultrasonic signal is released using loudspeaker；
1.2) signal sampling is carried out by Mike point；
1.3) frequency-amplitude vector x is obtained by Fast Fourier Transform (FFT)t', plus obtaining data matrix after time dimension
X'。
Described single-frequency ultrasonic signal frequency is 20KHz, sample frequency 44.1KHz.
Described step 2) specifically includes following steps：
2.1) decentralization is carried out to each data matrix X' and vectorization handles to obtain decentralization data vector xk, and
Integrate all m decentralization data vector xkObtain training data matrix X；
2.2) training data matrix X is decomposed into by singular value matrix Σ and two eigenvalue matrix U by singular value decomposition
And W, i.e. X=U Σ WT, what matrix W was is classified as the direction of singular value σ characteristic attribute；
2.3) the preceding d row for choosing matrix W form projection matrix W'=[w1,w2,…,wd]；
2.4) d principal character δ={ δ is obtained by formula δ=XW'1,δ2,….δd}。
Described d passes through formulaTo ask for, wherein：σiIt is big in training data matrix X i-th
Singular value, t for reconstruct threshold values.
Described multi-categorizer model θ includesIndividual two grader, each two grader are directed to a specific dangerous row
It is v for, its classification results, v ∈ { 0,1 }, multi-categorizer model θ classification resultsWherein：K is danger
Dangerous behavior species,Represent the ballot to all two graders classification results, the classification results of multi-categorizer model
C meets Vc(e)=k-1 is then corresponding hazardous act.
Described forest gradient former isWherein：τ1,τ2,…,τnFor a discrete gradient when
Between, obtained by the relation between the completeness α and duration τ of different hazardous acts,K ∈ 1,2 ..., and n } it is pair
K-th of the multi-categorizer model answered.
Described step 5) specifically includes following steps：
5.1) average energy E in sliding window and threshold value λ is carried out by window cutting algorithm on 20KHz Frequency points
Compare, data corresponding to E >=λ part are subjected to incremental cuts, be continuously available the action data X that the duration is ττ；
5.2) forest gradient former is substituted into, if the classification results of output, which are shown, belongs to dangerous driving behavior, for dangerous row
To send alarm.
The present invention relates to a kind of driver's hazardous act identifying system, including：Digital sampling and processing, principal character carry
Modulus block, forest gradient former training module and online hazardous act identification module, wherein：Digital sampling and processing and master
Want characteristic extracting module to be connected and be transferred through the data matrix that pretreatment obtains afterwards, principal character extraction module and forest ladder
Degree model training module is connected and transmits the principal character of extraction, and forest gradient former training module identifies with online hazardous act
Module is connected and transmits forest gradient former information, and data and forest of the online hazardous act identification module based on online acquisition are terraced
Degree model provides the recognition result of hazardous act in real time.
Brief description of the drawings
Fig. 1 is schematic flow sheet of the present invention；
Fig. 2 is four kinds of hazardous act schematic diagrames；
Fig. 3 is the frequency spectrum of four kinds of hazardous acts；
Fig. 4 is the behavior of leaning forward and data distribution comparison diagram of other hazardous acts on two-dimensional space；
Fig. 5 is the completeness and the fit correlation curve of duration of hazardous act；
Fig. 6 is the average energy value figure for the behavior of leaning forward；
Fig. 7 is accuracy rate, recall rate and the F value schematic diagrames of hazardous act identification；
Fig. 8 is the negative and positive class rate box-shaped figure of different hazardous acts；
Fig. 9 is the cumulative distribution rate schematic diagram of hazardous act recognition time；
Figure 10 is recognition accuracy schematic diagram；
Figure 11 is the accuracy rate schematic diagram under intelligent terminal diverse location.
Embodiment
Embodiments of the invention are elaborated below, the present embodiment is carried out lower premised on technical solution of the present invention
Implement, give detailed embodiment and specific operating process, but protection scope of the present invention is not limited to following implementation
Example.
Embodiment 1
As shown in figure 1, in the present embodiment, ultrasound field is set in vehicle first, obtains hazardous act in Doppler's frequency
Different mode in spectrum, principal character then is extracted through principal component analysis, different danger are identified by algorithm of support vector machine
The multi-categorizer of dangerous behavior, and gradient former forest is established according to the completeness of hazardous act and duration, finally will be real-time
Ultrasonic signal goes out hazardous act by gradient former Forest mapping after window cutting algorithm is cut into slices and given a warning.
As shown in Fig. 2 described typical hazard behavior includes leaning forward (Fetching Forward), feeding (Eating＆
Drinking), four classes such as rotary head (Turning Back) and lost article found (Picking up Drops).
Specifically include following steps：
1) receive in-car ultrasonic signal and handle, obtain the data matrix X' of ultrasonic signal spectrum structure.
1.1) single-frequency ultrasonic signal is released using loudspeaker.20KHz list is sent using the loudspeaker in intelligent terminal
Frequency ultrasonic signal.
1.2) signal sampling is carried out by Mike point.With 44.1KHz frequency samplings, information of the signal in time domain is obtained.
1.3) by Fast Fourier Transform (FFT) obtain frequency-amplitude vector x 't, plus obtaining data matrix after time dimension
X'。
As shown in figure 3,1024 dimension frequency-amplitudes vector is obtained for the Fast Fourier Transform (FFT) of 2048 points by window width
x′t, it represents the amplitude corresponding to the discrete frequency being evenly distributed of the t from 0Hz to 22.05KHz.Plus time dimension
Afterwards, obtain representing the data matrix X' of ultrasonic signal spectrum structure in the time of one end.
2) principal character δ={ δ is extracted by principal component analysis to some data matrix X'1,δ2,....δd}。
2.1) decentralization processing is carried out to each data matrix, obtains training data matrix X.To each data matrix X'
Decentralization processing is carried out, i.e.,By matrix X'kVectorization, it is launched into vector xk, synthesis instruction
Practice data matrix X.
2.2) training data matrix X is decomposed into by singular value matrix Σ and two eigenvalue matrix U by singular value decomposition
And W, i.e. X=U Σ WT, the row of matrix W are the direction of characteristic attribute.Singular value matrix Σ is diagonal matrix, element therein,
That is singular value σiBy arranging from big to small on leading diagonal, and the i-th of matrix W the row are singular value σiCorresponding characteristic attribute,
Singular value σiBigger, the characteristic attribute that homography W the i-th row represent is more important for training data matrix X.
2.3) the preceding d row for choosing matrix W form projection matrix W'=[w1,w2,…,wd]。
2.4) d principal character δ={ δ is obtained by δ=XW'1,δ2,....δd}。
In order to avoid there is over-fitting in learning process, it is necessary to cause under conditions of data message former enough is retained main
Number of features d is as small as possible, now can contemplate the reconstitution principle of principal component analysis, i.e.,Wherein：σi
For the i-th big singular value in training data matrix X, t is reconstruct threshold values.To ensure the validity of principal character, t takes 0.95, d to take
17.As shown in figure 4, when d takes 2, four hazardous acts are distributed in two-dimensional feature space shows only two principal characters of needs
Can causes between hazardous act two all right one way or the other points.
3) based on principal character and supporting vector machine model structure multi-categorizer model θ.
Two-dimentional principal character and supporting vector machine model based on extraction, between being acted to the typical hazardous act of each two
Training obtains two graders, available altogetherIndividual two grader, k is the species of typical hazard behavior, specific to some
Hazardous act classification throw 1, other classes throw 0.For a behavior e, its obtained all ticket isClassification
As a result it isClassification results also need to meet V simultaneouslyc(e)=k-1, now classification results c is pair
The c kind hazardous acts of result are answered, if being unsatisfactory for Vc(e)=k-1 is then divided into " other behaviors " class, and this is multi-categorizer model
θ。
4) it is gloomy to establish gradient former by the multiple multi-categorizer model θ of completeness α and duration τ integration for hazardous act
Woods.
Define one and start from t0End at t1Action e be in the completeness of tPin
Relation between the completeness α and duration τ of different hazardous acts is modeled, i.e. gradient former forest.
As shown in figure 5, the deadline T of every kind of different hazardous act is approximate to meet a Gaussian Profile, therefore to not
The deadline of same behavior is fitted with Gaussian Profile.In Gaussian Profile, the scope of neighbouring twice of the variances sigma of mean μ can be with
Cover more than 95% data, the present invention takes T=μ to each hazardous act, T=μ -2 σ and T=μ tri- distinguished points of+2 σ it is complete
The relation straight line between three completeness α and duration τ is obtained into the time.Afterwards a bit (T=is respectively taken on this three straight lines
Starting point corresponding to μ -2 σ, midpoint corresponding to T=μ, terminal corresponding to the σ of T=μ+2), it is fitted a quadratic function with these three points and makees
For the relation curve between completeness α and duration τ.
By duration it is discrete be n dimension gradient timetable vector T={ τ1,τ2,…,τn, according to this vector
And per the relation between class hazardous act completeness α and duration τ, cutting is carried out to training data matrix X, obtains gradient
Training dataIt is trained again by the above method, obtains gradient former forest
Described gradient former forestAs n multi-categorizer model θ integration.
5) by window cutting algorithm by duration τ action data XτGradient former forest is sent into, identifies danger
Behavior then sends alarm.
5.1) average energy E in sliding window and threshold value λ is carried out by window cutting algorithm on 20KHz Frequency points
Compare, data corresponding to E >=λ part are subjected to incremental cuts, be continuously available the action data X that the duration is ττ；
5.2) forest gradient former is substituted into, if the classification results of output, which are shown, belongs to dangerous driving behavior, for dangerous row
To send alarm.
As shown in fig. 6, when there is action to occur, the energy on Frequency point near 20kHz frequencies is due to Doppler effect
Influence substantially uprise.Therefore, the average of the energy on the Frequency point near 20kHz frequencies is taken to pass through cunning as basis for estimation
Dynamic window by the size of the energy value in window with threshold value compared with, so that it is determined that beginning that a driving behavior acts and knot
Beam.
It is determined that after the beginning of a behavior act, according to the time constantly by the data X of this actionτAnd current
Duration τ is incrementally sent into gradient former forest and is identified, until system provides recognition result or monitors that this is driven
Sail the end of behavior act.
It is τ ∈ [τ for a durationi,τi+1] action data section, system calls gradient former forest respectively
In i-th and i+1 sorter modelWithThis move is identified, only modelWithRecognition result
It is identical, just it is designated as once effectively identifying.Simultaneously, it is necessary to which the continuous effective identification for providing identical result for several times could export knot
Fruit, this continuous number are defined as confidence length l.And when system can not meet that confidence length can not provide recognition result,
System uses last grader in gradient former forestRecognition result as final result, now system model is degenerated
For the multi-categorizer model θ of gradient former forest is not used.
Driver's hazardous act identifying system in the present embodiment, including：Digital sampling and processing, principal character extraction
Module, forest gradient former training module and online hazardous act identification module, wherein：Digital sampling and processing with it is main
Characteristic extracting module is connected and is transferred through the data matrix that obtains afterwards of pretreatment, principal character extraction module and forest gradient
Model training module is connected and transmits the principal character of extraction, and forest gradient former training module identifies mould with online hazardous act
Block is connected and transmits forest gradient former information, data and forest gradient of the online hazardous act identification module based on online acquisition
Model provides the recognition result of hazardous act in real time.
When assessing driver's hazardous act recognition methods, the intelligence of the system of 8 driver driving installation loading present invention is eventually
The vehicle at end, is tested.
Assess index with being broadly divided into hazardous act action recognition accuracy rate and the knowledge of driver's hazardous act recognition methods
Other two aspects of time.The accuracy rate of identification is carved with F-score (F values) and False Positive Rate (negative and positive class rate)
Draw.F-score is an index for combining precision (precision) and recall (recall rate), can represent the accuracy of identification,
Formula isAnd what False Positive Rate were portrayed is the action for being not belonging to certain class
It is misidentified as the probability of such action.Recognition time can then be represented with CDF (cumulative distribution function) figures.
As shown in fig. 7, for whole test data set, for different typical hazard behavior acts, the present invention
Recognition accuracy be considerable.Specifically, all exist for all typical hazard behavior acts, the precision of identification
More than 89%, recall are both greater than 91%, and are all not less than 92% on most comprehensive index F-score.Show this
The system of invention ensure that basic recognition accuracy, be an actually available method.
As shown in figure 8, the driver for all 8 Data Collections that take one's test, it is shown that in identification process of the present invention
False Positive Rate.This index be for real system it is very important because system will as far as possible avoid by
Normal driving behavior is mistakenly identified as hazardous act and produces puzzlement to driver.As can be seen that system is endangered for different typical cases
Dangerous behavior act highest False Positive Rate are also not above 2.5%, and average value is even more low up to 1.4%, it is seen that is
System is very friendly for a user.
As shown in figure 9, for the synthesis of different typical hazard behavior acts and all typical hazard behavior acts, display
The cumulative distribution function of recognition time.As can be seen that 50% typical hazard behavior act can be identified in 1.4s
Come, and when this time reaches 2.3s, the typical hazard behavior act more than 80% can be recognized by the system.And typical hazard row
Average completion time for action is 4.6s.That is, the typical hazard behavior act more than 80% can be in completeness not
It is recognized by the system during to 50%, the EARLY RECOGNITION that the system of realizing is pursued.
As shown in Figure 10, for different traffic and condition of road surface, it is shown that system is for different typical hazards
The recognition accuracy of behavior act.As can be seen that the recognition accuracy of system is relatively stable, and with traffic and road like
The difference of condition is slightly different.Specifically, it is declined slightly in the peak period of traffic with respect to non-peak period recognition accuracy；Exist simultaneously
Ordinary road is also lower slightly relative to expressway discrimination.Show that system has preferably for different traffic and condition of road surface
Robustness.
As shown in figure 11, the different placement locations for intelligent terminal in vehicle, it is shown that system is for different allusion quotations
The recognition accuracy of type hazardous act action.As can be seen that the recognition accuracy of system is relatively stable, and as intelligent terminal is put
The difference of seated position slightly has difference.Best results near the guidance panel of vehicle are placed on, are placed on the left and right sides position of driver
Put effect to take second place, it is slightly worse to be placed on effect in the pocket of driver.But gap all very littles, also have in the case of worst 88% with
On accuracy of identification.Show that system is not strict with for the placement location of intelligent terminal in the car, system under each position
Can normal work, it is shown that the convenience and robustness of system.
Compared with prior art, the present invention can be in the early stage of hazardous act while accuracy of identification is ensured
Can recognition result go out and provide alarm, and the present invention using smart mobile phone sends ultrasonic signal, it is not necessary to relies on any
Additional external equipment, it is not easy to disturbed by external environments such as weather.
Claims (6)
1. a kind of driver's hazardous act recognition methods, it is characterised in that ultrasound field is set in vehicle first, obtains danger
Different mode of the behavior on Doppler frequency spectrum, principal character then is extracted through principal component analysis, passes through algorithm of support vector machine
The multi-categorizer of different hazardous acts is identified, and forest gradient-norm is established according to the completeness of hazardous act and duration
Type, real-time ultrasound ripple signal is finally identified into hazardous act after window cutting algorithm is cut into slices by forest gradient former
And give a warning；
Methods described specifically includes following steps：
1) receive in-car ultrasonic signal and handle, obtain the data matrix X' of ultrasonic signal spectrum structure；
2) d principal character δ={ δ is extracted by principal component analysis to some data matrix X'1,δ2,....δd}；
3) based on principal character δ and supporting vector machine model structure multi-categorizer model θ；
4) the completeness α and duration τ of hazardous act are directed to, forest gradient former Θ is established with reference to multi-categorizer model θ；
5) by window cutting algorithm by the action data X in duration ττForest gradient former Θ is sent into, identifies dangerous row
Then to send alarm, following steps are specifically included：
5.1) on 20KHz Frequency points by window cutting algorithm, by average energy E in sliding window compared with threshold value λ,
Data corresponding to E >=λ part are subjected to incremental cuts, are continuously available the action data X that the duration is ττ；
5.2) forest gradient former is substituted into, if the classification results of output, which are shown, belongs to dangerous driving behavior, is sent out for hazardous act
Go out alarm；
Described forest gradient former isWherein：τ1,τ2,...,τnFor a discrete gradient timetable,
Obtained by the relation between the completeness α and duration τ of different hazardous acts,As corresponding kth
Individual multi-categorizer model；
Described multi-categorizer model θ includesIndividual two grader, each two grader are directed to a specific hazardous act, its
Classification results are v, v ∈ { 0,1 }, multi-categorizer model θ classification resultsWherein：K is dangerous row
For species,The ballot to all two graders classification results is represented, the classification results c of multi-categorizer model expires
Sufficient Vc(e)=k-1 is then corresponding hazardous act.
2. driver's hazardous act recognition methods according to claim 1, it is characterized in that, described step 1) specifically includes
Following steps：
1.1) single-frequency ultrasonic signal is released using loudspeaker；
1.2) signal sampling is carried out by Mike point；
1.3) by Fast Fourier Transform (FFT) obtain frequency-amplitude vector x 't, plus obtaining data matrix X' after time dimension.
3. driver's hazardous act recognition methods according to claim 2, it is characterized in that, described single-frequency ultrasonic signal
Frequency is 20KHz, sample frequency 44.1KHz.
4. driver's hazardous act recognition methods according to claim 1, it is characterized in that, described step 2) specifically includes
Following steps：
2.1) decentralization is carried out to each data matrix X' and vectorization handles to obtain decentralization data vector xk, and integrate
All m decentralization data vector xkObtain training data matrix X；
2.2) training data matrix X is decomposed into by singular value matrix Σ and two eigenvalue matrix U and W by singular value decomposition,
That is X=U Σ WT, the direction of the characteristic attribute for being classified as singular value σ of matrix W；
2.3) the preceding d row for choosing matrix W form projection matrix W'=[w1,w2,...,wd]；
2.4) d principal character δ={ δ is obtained by formula δ=XW'1,δ2,....δd}。
5. driver's hazardous act recognition methods according to claim 4, it is characterized in that, described d passes through formulaTo ask for, wherein：σ i are the i-th big singular value in training data matrix X, and t is reconstruct threshold values.
A kind of 6. driver's hazardous act identifying system for realizing claim 1 methods described, it is characterised in that including：Data
Acquisition processing module, principal character extraction module, forest gradient former training module and online hazardous act identification module, its
In：Digital sampling and processing is connected with principal character extraction module and is transferred through the data matrix that obtains afterwards of pretreatment,
Principal character extraction module is connected with forest gradient former training module and transmits the principal character of extraction, and forest gradient former is instructed
Practice module with online hazardous act identification module to be connected and transmit forest gradient former information, online hazardous act identification module base
The recognition result of hazardous act is provided in real time in the data and forest gradient former of online acquisition.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201611116789.2A CN106650644B (en) | 2016-12-07 | 2016-12-07 | The recognition methods of driver's hazardous act and system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201611116789.2A CN106650644B (en) | 2016-12-07 | 2016-12-07 | The recognition methods of driver's hazardous act and system |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN106650644A CN106650644A (en) | 2017-05-10 |
| CN106650644B true CN106650644B (en) | 2017-12-19 |
Family
ID=58818715
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201611116789.2A Active CN106650644B (en) | 2016-12-07 | 2016-12-07 | The recognition methods of driver's hazardous act and system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106650644B (en) |
Families Citing this family (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107797681B (en) * | 2017-10-13 | 2019-03-15 | 上海交通大学 | Steering wheel for vehicle rotation method for tracing and system based on intelligent mobile terminal audio devices |
| CN109063751B (en) * | 2018-07-16 | 2021-09-17 | 江苏智通交通科技有限公司 | Traffic high-risk personnel identification method based on gradient lifting decision tree algorithm |
| CN111055849B (en) * | 2018-10-17 | 2021-04-06 | 财团法人车辆研究测试中心 | Intersection intelligent driving method and system based on support vector machine |
| CN110596490A (en) * | 2019-09-11 | 2019-12-20 | 天津科技大学 | Intelligent detection method for railway turnout fault |
| CN113593183A (en) * | 2021-06-25 | 2021-11-02 | 湖南大学 | Detection method, device, equipment and medium for fatigue driving and distraction driving based on acoustics |
Family Cites Families (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US4259665A (en) * | 1979-05-29 | 1981-03-31 | Rmr Systems, Inc. | Driver sleep or fatigue alarm |
| CN103347446B (en) * | 2010-12-10 | 2016-10-26 | Tk控股公司 | For monitoring the system of vehicle driver |
| CN102106730A (en) * | 2011-03-16 | 2011-06-29 | 上海交通大学 | Method for processing electroencephalogram signal and detecting alertness based on fractal characteristics |
| CN102567743A (en) * | 2011-12-20 | 2012-07-11 | 东南大学 | Automatic identification method of driver gestures based on video images |
-
2016
- 2016-12-07 CN CN201611116789.2A patent/CN106650644B/en active Active
Also Published As
| Publication number | Publication date |
|---|---|
| CN106650644A (en) | 2017-05-10 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106650644B (en) | The recognition methods of driver's hazardous act and system | |
| CN108596221B (en) | Image recognition method and device for scale reading | |
| CN111428631B (en) | Visual identification and sorting method for unmanned aerial vehicle flight control signals | |
| CN109272017B (en) | Vibration signal mode identification method and system of distributed optical fiber sensor | |
| CN107784276B (en) | Microseismic event identification method and device | |
| CN103780899A (en) | Method and device for detecting whether camera is interfered and video monitoring system | |
| CN105718912B (en) | A kind of vehicle characteristics object detecting method based on deep learning | |
| CN101987017A (en) | Electroencephalo-graph (EEG) signal identification and detection method for measuring alertness of driver | |
| CN109543647A (en) | A kind of road abnormality recognition method, device, equipment and medium | |
| CN104268598A (en) | Human leg detection method based on two-dimensional scanning lasers | |
| CN109284705A (en) | A kind of method for detecting abnormality of Traffic Surveillance Video | |
| CN111738218A (en) | Human body abnormal behavior recognition system and method | |
| CN116259002A (en) | Human body dangerous behavior analysis method based on video | |
| CN106708009A (en) | Ship dynamic positioning measurement system multiple-fault diagnosis method based on support vector machine clustering | |
| Zhao et al. | Detection of road surface anomaly using distributed fiber optic sensing | |
| Pontoppidan et al. | Condition monitoring with mean field independent components analysis | |
| KR20190050551A (en) | Apparatus and method for recognizing body motion based on depth map information | |
| CN109409216B (en) | Speed self-adaptive indoor human body detection method based on subcarrier dynamic selection | |
| KR101300650B1 (en) | A method of recognizing pri modulation type based on support vector machine, and an apparatus of recognizing pri modulation type | |
| CN109492573A (en) | A kind of pointer read method and device | |
| WO2008055306A1 (en) | Machine learning system for graffiti deterrence | |
| CN116486177A (en) | Underwater target identification and classification method based on deep learning | |
| CN113887634B (en) | Electric safety belt detection and early warning method based on improved two-step detection | |
| Cvijetić et al. | Deep learning-based vehicle speed estimation using the YOLO detector and 1D-CNN | |
| CN116612638A (en) | Traffic collision accident detection method, device and readable medium |
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 |