CN105496369A - Angry driving state detection method - Google Patents

Abstract

The invention discloses an angry driving state detection method. To solve deficiencies of poor detection accuracy, high hardware cost, uneasy wearing of equipment and susceptibleness to environment factors of an angry state detection mechanisms in the prior art, through observation of pulse change conditions of a driver under a normal and angry driving states, and related comparison experiments, three features related to the angry mood are determined: pulse amplitude change rate, decreasing amplitude from a dominate wave to a dicrotic wave and a change rate of pulse rate. Based on the feature data, a learning assessment model for driver angry driving state is formed; the model comprises modules of pulse data collection and preprocessing, feature extraction, feature training and state adjusting; and angry driving state detection can be effectively achieved via the learning assessment model. High angry driving state identification rate can be acquired with low hardware cost by the use of the angry driving state detection method.

Description

A kind of angry driving condition detection method

Technical field

The present invention relates to driving condition detection method, particularly relate to a kind of angry driving condition detection method, belong to the interleaving techniques application of mobile computing, sensing technology and signal processing.

Background technology

Aggressive driving refers to by the increasing the weight of and produced potential impact of driving pressure psychologically, deliberately other people in driving environment is carried out to the behavior of health, psychology or emotion injury, is divided into low level, medium rank and high-level three kinds of attack ranks.Comprise and deliberately knock into the back, rob and make a dash across the red light, have a mind to intercept, furious driving, sudden strain of a muscle headlight, to blow a whistle loudly, Aggressive driving has huge harm to road traffic.The angry driving condition of driver is the major reason producing Aggressive driving, once be in angry driving condition, the words and deeds of driver will be difficult to self-control.Indignation is driven the vehicle accident caused and is happened occasionally, serious harm traffic safety.The year junior one on February 10 in 2013, on Beijing-Shanghai Expressway, Liu's drive private car is travelled outdoors, and meet front truck and take fast always, blowing a whistle several times has no resolution for surmounting front truck, result rage, then opens the throttle to surmount on the right side of front truck; Front truck is unwilling to be outshone, and also accelerates to surmount Liu's car, and sail the high speed of competing with one another is competing, car to car impact, causes the major traffic accidents that Liu is dead, front truck driver is injured.

In time, the angry driving condition of driver is detected rapidly and to be regulated in time be the key of the vehicle accident avoiding causing due to Aggressive driving.The testing mechanism of current angry driving condition has had a series of achievement in research.Research worker is had to propose based on continuous facial image indignation driving mood testing mechanism, apply the facial expression change of mutual subspace method identification driver, and then the rank of the angry driving mood of assess driver, but these class methods are confined to the good situation of driver's cabin light, the situation of driving at night cannot be applied to, have larger limitation.Aggressive driving is detected in relevant numerous parameters, has research worker to propose to select most suitable parameter to detect the Aggressive driving of driver, effectively reduce the Space-time Complexity of detection system.Research worker is had to propose a kind of aggressive driving state detection algorithm based on driving event, based on seasonal effect in time series glossary of symbols approximate data (SymbolicAggregateApproximation, SAX), analytical data from the sensor of smart mobile phone, with assembly time sequence data effectively, the Detection accuracy of algorithm reaches 50% ~ 66.7%.Driving behavior evaluating system JoinDriving based on smart mobile phone can help driver to notice, and whether the driving behavior of oneself has aggressivity, by drive check and evaluation module and based on driving event analytical model analyze acceleration transducer, direction sensor and GPS sensor acquisition to data, whether the driving of then applying marking system assess driver has aggressivity, these class methods belong to indirect detection strategy, easily cause larger metrical error.Method hardware cost based on the angry emotion of multiple index analysis people such as blood pressure, electroencephalogram, cardiac impedance, facial electromyography is higher, and wears not easily.Also has the impact of research worker research driver individual factors driver indignation driving condition, in the driving task of special time particular course, real time record driver indignation grade, finds that sex, age and different driving style have larger impact to driver's indignation emotion.In a word, the achievement in research ubiquity of current angry driving condition testing mechanism that Detection accuracy is not high, hardware cost is higher, equipment is worn not easily, by defects such as such environmental effects are larger.

For the problems referred to above, the present invention adopts the pulse transducer collection user pulse signal being easy to wear, the change of pulse frequency that driver's indignation driving condition is correlated with, these three features of fall between the amplitude of main ripple and main ripple to dicrotic wave are extracted by contrast experiment, train the learning evaluation model of angry driving condition, propose a kind of angry driving condition detection method, higher angry driving condition discrimination can be obtained in the situation of low hardware cost.

Summary of the invention

Technical problem to be solved by this invention provides a kind of angry driving condition detection method for the deficiency of background technology.

The present invention is for solving the problems of the technologies described above by the following technical solutions

A kind of angry driving condition detection method, specifically comprises following steps:

Step 1, gathers the pulse data parameter of driver under angry state in advance;

Step 2, carries out feature extraction to the pulse data parameter that step 1 obtains;

Step 3, is carried out features training to the characteristic parameter that step 2 is extracted, utilizes the study mechanism that signal forward-propagating and error oppositely regulate, learnt and then build learning evaluation model by successive ignition;

Step 4, Real-time Collection driver pulse data parameter;

Step 5, the learning evaluation model that the driver's pulse data parameters input step 3 step 4 gathered is set up, and then show whether driver is in angry driving condition, if angry driving condition, driver is then reminded to carry out state adjustment, otherwise, then return step 4.

As the further preferred version of a kind of angry driving condition detection method of the present invention, feature extraction described in step 2 comprises the rate of change, the range of decrease of main ripple to dicrotic wave and the rate of change of pulse frequency that extract pulse amplitude.

As the further preferred version of a kind of angry driving condition detection method of the present invention, described learning evaluation model comprises input layer, hidden layer and output layer.

As the further preferred version of a kind of angry driving condition detection method of the present invention, described step 3 specifically comprises following steps:

Step 3.1: the characteristic parameter that input step 2 extracts: the rate of change of pulse amplitude, the range of decrease of main ripple to dicrotic wave and the rate of change of pulse frequency;

Step 3.2: calculate input layer respectively to the output of hidden layer and hidden layer to the output of output layer according to the characteristic parameter received;

Step 3.3: according to $V_k=\frac{1}{2}\underset{j}{\&Sigma;}{(O_{kj}-^O_{kj})}^2<{\mathrm{\&delta;}}_{ANN},V=\underset{k}{\&Sigma;}{V}_k\&le;{\mathrm{\&delta;}}_{ANN}$ Calculate output layer respectively to each layer component output error of hidden layer and hidden layer to the output error of output layer, if component error used meets error requirements, then learning process terminates, otherwise enters step 3.4; Wherein, δ _aNNrepresentative errors threshold value, V _kthe error that representative uses k sample learning to produce, O _kjrepresent the actual Output rusults of a kth sample learning of jth layer network, ^O _kjrepresent the desired output result of a kth sample of jth layer network, the total error of V representative sample collection;

Step 3.4: according to ${\mathrm{\&Delta;W}}_{ij+1}={\mathrm{a\&phi;}}_i{O}_j+\&part;{\mathrm{\&Delta;W}}_{ij},O_j=f\left(\underset{i=0}{\overset{N-1}{\&Sigma;}}{W}_{ij}{X}_i-{\mathrm{\&theta;}}_1\right)$ Calculate the error of output layer, hidden layer and input layer successively, wherein: Δ W _ij+1represent the correction of i-th sample at the connection weights of jth+1 layer network, a represents Studying factors, φ _irepresent the error of output node, O _jrepresent the output valve of a hidden layer jth node, represent modifying factor, X _irepresent the output of output layer i-th node, W _ijrepresent the weight of a hidden layer jth node and output layer i-th node, O _jthe threshold value of a hidden layer jth node, j=0,1 ..., M-1, N represent the quantity of input layer, and M represents the quantity of hidden layer;

Step 3.5: according to the error of output layer, hidden layer and input layer that step 3.4 calculates, and then adjust output layer, connection weights between hidden layer and input layer successively; And then to be learnt by successive ignition and then to build learning evaluation model, until all training samples are all trained complete.

As the further preferred version of a kind of angry driving condition detection method of the present invention, in step 3.4, described modifying factor value is between 0 to 1.

The present invention adopts above technical scheme compared with prior art, has following technique effect:

1, the present invention's indignation state-detection accuracy rate is higher, experiment shows the accuracy rate that can reach 66.7%: driver is actual be in not angry driving condition time, systems axiol-ogy accuracy rate is 83.33%, and driver is actual when being in angry driving condition, and systems axiol-ogy accuracy rate is 50%;

2, present system response time is short, and driver's indignation driving condition detection system has the delay of about 1s, and major expenses is on pulse data sample and transform;

3, Space-time Complexity of the present invention is low, and the time complexity of method is O (n); Installed System Memory consumes, and Installed System Memory consumption is between 50-60M.

Accompanying drawing explanation

Fig. 1 is angry driving condition detection model figure of the present invention;

Fig. 2 is Applied Learning assessment models driver of the present invention indignation driving condition testing process.

Detailed description of the invention

Below in conjunction with accompanying drawing, technical scheme of the present invention is described in further detail:

As shown in Figure 1, a kind of angry driving condition detection method, specifically comprises following steps:

Step 1, gathers the pulse data parameter of driver under angry state in advance;

Step 2, carries out feature extraction to the pulse data parameter that step 1 obtains;

Step 3, is carried out features training to the characteristic parameter that step 2 is extracted, utilizes the study mechanism that signal forward-propagating and error oppositely regulate, learnt and then build learning evaluation model by successive ignition;

Step 3.1: the characteristic parameter that input step 2 extracts: the rate of change of pulse amplitude, the range of decrease of main ripple to dicrotic wave and the rate of change of pulse frequency;

Step 3.2: calculate input layer respectively to the output of hidden layer and hidden layer to the output of output layer according to the characteristic parameter received;

Step 3.3: according to $V_k=\frac{1}{2}\underset{j}{\&Sigma;}{(O_{kj}-^O_{kj})}^2<{\mathrm{\&delta;}}_{ANN},V=\underset{k}{\&Sigma;}{V}_k\&le;{\mathrm{\&delta;}}_{ANN}$ Calculate output layer respectively to each layer component output error of hidden layer and hidden layer to the output error of output layer, if component error used meets error requirements, then learning process terminates, otherwise enters step 3.4; Wherein, δ _aNNrepresentative errors threshold value, V _kthe error that representative uses k sample learning to produce, O _kjrepresent the actual Output rusults of a kth sample learning of jth layer network, ^O _kjrepresent the desired output result of a kth sample of jth layer network, the total error of V representative sample collection;

Step 3.4: according to ${\mathrm{\&Delta;W}}_{ij+1}={\mathrm{a\&phi;}}_i{O}_j+\&part;{\mathrm{\&Delta;W}}_{ij},O_j=f\left(\underset{i=0}{\overset{N-1}{\&Sigma;}}{W}_{ij}{X}_i-{\mathrm{\&theta;}}_1\right)$ Calculate the error of output layer, hidden layer and input layer successively, wherein: Δ W _ij+1represent the correction of i-th sample at the connection weights of jth+1 layer network, a represents Studying factors, φ _irepresent the error of output node, O _jrepresent the output valve of a hidden layer jth node, represent modifying factor, X _irepresent the output of output layer i-th node, W _ijrepresent the weight of a hidden layer jth node and output layer i-th node, O _jthe threshold value of a hidden layer jth node, j=0,1 ..., M-1, N represent the quantity of input layer, and M represents the quantity of hidden layer;

Step 3.5: according to the error of output layer, hidden layer and input layer that step 3.5 calculates, and then adjust output layer, connection weights between hidden layer and input layer successively; And then to be learnt by successive ignition and then to build learning evaluation model, until all training samples are all trained complete;

Step 3.5: according to the error of output layer, hidden layer and input layer that step 3.5 calculates, and then adjust output layer, connection weights between hidden layer and input layer successively; And then to be learnt by successive ignition and then to build learning evaluation model, until all training samples are all trained complete;

Step 4, Real-time Collection driver pulse data parameter;

Step 5, the learning evaluation model that the driver's pulse data parameters input step 3 step 4 gathered is set up, and then show whether driver is in angry driving condition, if angry driving condition, driver is then reminded to carry out state adjustment, otherwise, then return step 4.

Wherein, described feature comprises the rate of change of pulse amplitude, the range of decrease of main ripple to dicrotic wave and the rate of change of pulse frequency, and described learning evaluation model comprises input layer, hidden layer and output layer.

The angry driving condition detection method that the present invention proposes is divided into off-line and online two stages: the data source of off-line phase is the pulse data of driver under angry state of pre-acquired, ultimate principle carries out feature extraction to these pulse data, these features are carried out learning training, obtains judging whether driver is in the learning evaluation model of angry driving condition; The data source of on-line stage is the pulse data that Real-time Collection arrives, and carries out feature extraction equally to these data, and after feature extraction terminates, application off-line phase trains the learning evaluation model inspection driver obtained whether to be in angry driving condition.

Indignation driving condition detection model: angry driving condition detection model comprises the module such as pulse data collection and pretreatment, feature extraction, features training, status adjustment as shown in Figure 1: the pulse data that will gather driver in multiple angry driving procedure during the operation prerequisite of angry driving condition detection system in advance, then go out the feature relevant with human body indignation state of our needs from these extracting data, train corresponding angry driving condition detected rule.The detection of indignation driving condition is the implementation of an order, indignation driving condition detection module calls the pulse data of acquisition, then feature extraction is carried out to these data, extract the feature relevant with driver's indignation driving condition, application precondition good learning evaluation model judges whether driver is in angry driving condition, finally can utilize the driving condition of corresponding stimulation means adjustment driver according to testing result.

Feature extraction is by the incoherent characteristic information of initial data filtering of complexity, extracts useful characteristic parameter construction feature space.The present invention is based on statistical analysis technique and by concrete contrast experiment, find the feature relevant to driver's indignation state to set up effective learning evaluation model.Having organized 12 experimenters altogether, is all healthy adults, and the age is between 22 to 26 years old, and pulse value data when experimenter being in normal emotional state and being in angry state are recorded, as shown in table 1.

Table 1

In Table 1, the main wave amplitude of pulse is by system acquisition gained, and the main ripple range of decrease and pulse rate calculations gained, wherein the main ripple range of decrease is the change of change than upper transverse axis amplitude of longitudinal axis amplitude in a bit of time, and pulse frequency is the mathematic interpolation gained according to putting between main ripple.As shown in Table 1, under angry state, main wave amplitude strengthens generally more than 20%, and the range of decrease of main ripple obviously becomes large, and pulse frequency increases about 10% simultaneously.Thus, the present invention has extracted the rate of change of pulse amplitude, main ripple to the rate of change three of the range of decrease of dicrotic wave and the pulse frequency feature relevant to driver's indignation state.

Set up the learning evaluation model of angry driving condition: the study mechanism that the learning evaluation model of angry driving condition mainly utilizes signal forward-propagating and error oppositely to regulate, learnt by successive ignition, build the intelligent network model of process nonlinear transformations, comprise input layer, hidden layer and output layer.

As mentioned above, the present invention selects main wave amplitude, main ripple to these three characteristic parameters of change of the range of decrease of dicrotic wave and pulse frequency as the input of assess driver indignation state, and namely input layer has three processing elements.Whether driver is in angry state as output level, 0 represents driver is not in angry driving condition, and 1 represents driver is in angry driving condition.The number of hidden layer processing elements is determined by empirical equation:

$n_h=\sqrt{n+m}+a---\left(1\right)$

In formula (1), n _hrepresent the number of hidden layer processing elements, n represents the number of input layer processing elements, and m represents the number of output layer processing elements, and a is a constant, specifically sets according to practical situation.

In the forward-propagating stage, input information is transmitted to output layer from input layer through hidden layer, produces output signal at outfan.In the transmittance process forward of signal, the weights of network immobilize, and the state of every one deck processing elements only affects the state of lower one deck processing elements.If the output expected can not be obtained at output layer, then proceed to error signal back-propagation process.In back-propagation phase, fail to meet the error signal of required precision by outfan, in some way successively forward direction, and error distribution is given all unit of each layer, adjust the connection weight of each elementary layer according to error signal dynamically.By the forward-propagating that goes round and begins again with oppositely regulate, the weights between processing elements are constantly revised.When the error outputed signal meets required precision, stop study.

The present invention adopts Sigmoid type non-linear threshold to set the threshold value of node, and activation primitive is as shown in formula (2):

$f\left(x\right)=\frac{1}{1+e^{-x}}---\left(2\right)$

Carry out sample learning process continuously, after setting up formula (3), (4), terminate learning process:

$V_k=\frac{1}{2}\underset{j}{\&Sigma;}{(O_{kj}-^O_{kj})}^2<{\mathrm{\&delta;}}_{ANN}---\left(3\right)$

$V=\underset{k}{\&Sigma;}{V}_k\&le;{\mathrm{\&delta;}}_{ANN}---\left(4\right)$

In formula (3), (4), δ _aNNrepresentative errors threshold value, V _kthe error that representative uses k sample learning to produce, O _kjrepresent the actual Output rusults of a kth sample learning of jth layer network, ^O _kjrepresent the desired output result of a kth sample of jth layer network, the total error of V representative sample collection;

If formula (3) does not satisfy condition, error feedback mechanism along initial path back propagation, can adjust weight and the threshold value of each layer network according to gradient descent method.Relatively Practical Calculation Output rusults and desired output result, can reduce training error by result of calculation adjustment threshold value and weight.Self study regulating error model is:

${\mathrm{\&Delta;W}}_{ij+1}={\mathrm{a\&phi;}}_i{O}_j+\&part;{\mathrm{\&Delta;W}}_{ij}---\left(5\right)$

In formula (5), Δ W _ij+1represent the correction of i-th sample at the connection weights of jth+1 layer network, a represents Studying factors, φ _irepresent the error of output node, O _jrepresent the output valve of a hidden layer jth node, represent modifying factor value usually between 0 to 1, O _jcalculating formula be:

$O_j=f\left(\underset{i=0}{\overset{N-1}{\&Sigma;}}{W}_{ij}{X}_i-{\mathrm{\&theta;}}_1\right),j=0,1,\mathrm{...},M-1---\left(6\right)$

In formula (6), X _irepresent the output of output layer i-th node, W _ijrepresent the weight of a hidden layer jth node and output layer i-th node, θ _jthe threshold value of a hidden layer jth node; N represents the quantity of input layer, and M represents the quantity of hidden layer.The calculating formula of input layer is same as described above.

The concrete steps of learning process are as follows:

First each is connected weights and be set to less random number, different weights are arranged for different processing elements;

Get out training sample, input parameter comprises the rate of change of pulse amplitude, the range of decrease of main ripple to dicrotic wave and the rate of change of pulse frequency.It is right that a training having neither part nor lot in training is selected in output at random from preprepared training sample, as the input amendment of input layer;

Input layer is calculated respectively to the output of hidden layer and hidden layer to the output of output layer according to above-mentioned formula;

Calculate output layer to each layer component output error of hidden layer and hidden layer to the output error of output layer, if component error used meets error requirements, then learning process terminates; Otherwise enter step 3.4;

Calculate the error of output layer, hidden layer and input layer successively;

Adjust output layer, connection weights between hidden layer and input layer successively;

Turn back to step 3.2, carry out the training of next round, until all training samples are all trained complete.

As shown in Figure 2: driver's indignation driving condition detects: after obtaining the learning evaluation model about driver's indignation driving condition, just can utilize the driving condition of this model evaluation driver.It is the time-domain value that the real-time pulse of driver changes that pulse transducer collects, and eigenvalue carries out feature extraction based on time-domain value.After obtaining these eigenvalues, Applied Learning assessment models detects driver and whether is in angry driving condition.Applied Learning assessment models detects the process of driver's indignation driving condition.

The pulse crest variation characteristic extracted and pulse frequency variation characteristic are all changing values, first need to extract respectively driver's pulse wave peak value in normal state and pulse frequency, the pulse time-domain value of Real-time Collection driver, the real-time pulse crest of driver and pulse frequency can be obtained, calculate the change of pulse wave peak value and the change of pulse frequency; The fall off rate of crest to dicrotic pulse prewave can be obtained from pulse time-domain value simultaneously; Using these three kinds of eigenvalues as the input of learning evaluation model, whether be in angry driving condition by learning evaluation model evaluation driver.Learning evaluation model is by training, and the weight coefficient connecting input layer and hidden layer and be connected hidden layer and output layer is stablized, and namely meets the error requirements of system.The process of application quantizes three eigenvalues exactly, and the standard of quantification and the input of training are consistent, and so just can obtain the Output rusults determined, and then can judge whether driver is in angry driving condition.

Claims (5)

1. an angry driving condition detection method, is characterized in that: specifically comprise following steps:

Step 1, gathers the pulse data parameter of driver under angry state in advance;

Step 2, carries out feature extraction to the pulse data parameter that step 1 obtains;

Step 3, is carried out features training to the characteristic parameter that step 2 is extracted, utilizes the study mechanism that signal forward-propagating and error oppositely regulate, learnt and then build learning evaluation model by successive ignition;

Step 4, Real-time Collection driver pulse data parameter;

Step 5, the learning evaluation model that the driver's pulse data parameters input step 3 step 4 gathered is set up, and then show whether driver is in angry driving condition, if angry driving condition, driver is then reminded to carry out state adjustment, otherwise, then return step 4.

2. the angry driving condition detection method of one according to claim 1, is characterized in that: feature extraction described in step 2 comprises the rate of change, the range of decrease of main ripple to dicrotic wave and the rate of change of pulse frequency that extract pulse amplitude.

3. the angry driving condition detection method of one according to claim 1, is characterized in that: described learning evaluation model comprises input layer, hidden layer and output layer.

4. the angry driving condition detection method of one according to claim 1, is characterized in that: described step 3 specifically comprises following steps:

Step 3.1: the characteristic parameter that input step 2 extracts: the rate of change of pulse amplitude, the range of decrease of main ripple to dicrotic wave and the rate of change of pulse frequency;

Step 3.2: calculate input layer respectively to the output of hidden layer and hidden layer to the output of output layer according to the characteristic parameter received;

Step 3.3: according to $V_k=\frac{1}{2}\underset{j}{\&Sigma;}{(O_{kj}-^O_{kj})}^2<{\mathrm{\&delta;}}_{ANN},V=\underset{k}{\&Sigma;}{V}_k\&le;{\mathrm{\&delta;}}_{ANN}$ Calculate output layer respectively to each layer component output error of hidden layer and hidden layer to the output error of output layer, if component error used meets error requirements, then learning process terminates, otherwise enters step 3.4; Wherein, δ _aNNrepresentative errors threshold value, V _kthe error that representative uses k sample learning to produce, O _kjrepresent the actual Output rusults of a kth sample learning of jth layer network, ^O _kjrepresent the desired output result of a kth sample of jth layer network, the total error of V representative sample collection;

Step 3.4: according to ${\mathrm{\&Delta;W}}_{ij+1}={\mathrm{a\&phi;}}_i{O}_j+\&part;{\mathrm{\&Delta;W}}_{ij},O_j=f\left(\underset{i=0}{\overset{N-1}{\&Sigma;}}{W}_{ij}{X}_i-{\mathrm{\&theta;}}_j\right)$ Calculate the error of output layer, hidden layer and input layer successively, wherein: Δ W _ij+1represent the correction of i-th sample at the connection weights of jth+1 layer network, a represents Studying factors, φ _irepresent the error of output node, O _jrepresent the output valve of a hidden layer jth node, represent modifying factor, X _irepresent the output of output layer i-th node, W _ijrepresent the weight of a hidden layer jth node and output layer i-th node, O _jthe threshold value of a hidden layer jth node, j=0,1 ..., M-1, N represent the quantity of input layer, and M represents the quantity of hidden layer;

Step 3.5: according to the error of output layer, hidden layer and input layer that step 3.4 calculates, and then adjust output layer, connection weights between hidden layer and input layer successively; And then to be learnt by successive ignition and then to build learning evaluation model, until all training samples are all trained complete.

5. the angry driving condition detection method of one according to claim 4, is characterized in that: in step 3.4, described modifying factor value is between 0 to 1.