CN103729626A - Human body heat source feature extracting and distinguishing method based on infrared pyroelectric information - Google Patents

Abstract

The invention belongs to the technical field of human body distinguishing and discloses a human body heat source feature extracting and distinguishing method based on infrared pyroelectric information. The method aims to achieve 360-degree wide-range and remote detection and detection of static infrared heat sources and effectively solves the problems of confliction between comfort and energy saving of an intelligent air-conditioner and high false alarm rate of a safety supervision system. According to the technical scheme, the method comprises the steps that a stepping motor is adopted to drive a single pyroelectric infrared detector to rotate at a uniform speed, and then remote, 360-degree wide-range and static object detection is achieved; the detection range is a disk shape, wavelet packet analysis is conducted on collected human body heat source samples and non-human-body heat source samples, a signal wavelet entropy is taken as the features of the signal, 5-fold cross-validation is conducted by means of the BP neural network, and then human body heat resources are distinguished from non-human-body heat resources. The method is mainly applied to human body distinguishing.

Description

The feature extraction of human body thermal source and method of discrimination based on infrared thermal release electric information

Technical field

The invention belongs to human body discrimination technology field.Utilize pyroelectric infrared detector to carry out human body and non-human thermal source is sentenced method for distinguishing.Specifically, relate to the feature extraction of human body thermal source and the method for discrimination based on infrared thermal release electric information.

Background technology

The research of human body discrimination technology has great using value and Research Significance.On the one hand: people are more and more interested in Smart Home in recent years, and the control product of Smart Home some become household electrical appliance, some is just becoming household electrical appliance.Air-conditioning, as a part indispensable in household electrical appliances, nearly ten years, has experienced the stage of high speed development.Comfortableness and the energy-conservation important indicator that all becomes this class luxury goods of air-conditioning.By the detection of position of human body and activity is carried out to intellectuality to air-conditioning, control, reduce unnecessary temperature and regulate, can meet comfortable and energy-conservation dual indexes.Yet the interference of non-human infrared origin makes people's health check-up information measure existing error-detecting, and this will produce unnecessary waste of energy to the use of intelligent air condition.On the other hand: people have proposed safely higher technical requirement to social public security and domestic environment, as one of modal monitoring product in intrusion alarm system, there is the shortcoming of high rate of false alarm in rpyroelectric infrared (PIR) detector.Therefore studying a kind of human body recognition method based on rpyroelectric infrared signal is the key that solving error detects, and is also the effective means that reduces waste of energy.

The detection of human body target can mainly be divided into two classes: video camera detects and infrared sensor detects.Although it is high that video camera detects with tracking accuracy, data processing complex is expensive, occupies larger memory headroom and invasion of privacy, and its use is very restricted.Infrared sensor is also applicable to the detection of human body information very much.It is low that pyroelectric infrared sensor has price, and low-power is contactless, and good concealment is applied to the detection of human body information in intelligent environment widely to the advantage of illumination condition no requirement (NR).Pyroelectric infrared sensor can detect with non-contacting form the infrared energy variation of human body radiation, and converts thereof into voltage signal output.It is connected in series by 2 opposite polarity sensing elements conventionally, and fits together with 1 high resistant and 1 field effect transistor.Because 2 sensing element polarity are contrary, connection signal will be cancelled out each other and not output.Therefore pyroelectric sensor is to the variation of environment temperature, background radiation, and the random noise of self temperature variation and vibrated generation all has good compensating action.But it is only to the movement of human body or motion sensitive, and detection range is shorter.

The mid-90 in last century, the processing research of the human body information sensing based on pir sensor is devoted at the human environment system development center of Japanese Matsushita Electric Industrial company always, and achievement in research is applied with in Smart Home, comprises the control of air-conditioning, illumination etc.Nineteen ninety-five Nobuyuki Yoshiike etc. are the situation for detection of indoor human body by human information sensing system, personnel amount, position and active situation.Personnel amount recognition accuracy is 90%.This research and development centre improved system algorithm in 1997, and during the quantity at the gate that is applied to come in and go out detects, counting rate of accuracy reached to 98%.1998 Nian Gai centers have increased again by 4 channel distance sensors on original system basis, and are applied in indoor occupant quantity and position probing, the different gestures of human body can be detected, and the number recognition accuracy of system is 93%.Within 2006, Busan, Korea national university has researched and developed the indoor detection system based on PIR detector array.

Yet the interference of non-human infrared origin makes human body information occur error-detecting, this will produce unnecessary waste of energy to the use of intelligent air condition, also can bring the wrong report of intrusion alarm system.Studying a kind of human body recognition method based on rpyroelectric infrared signal is the key that solving error detects.Because the thermal source of different shapes is when by PIR sensing range, will produce different signals, this is just for utilizing pyroelectric sensor to carry out human body and non-human detection provides theoretic foundation.Although PIR detector has been obtained some achievements in the detection of human body and non-human thermal source, can't realize on a large scale simultaneously, the detection of static infrared heat source, and reach the higher target of single feature discrimination.

Summary of the invention

The present invention is intended to solution and overcomes the deficiencies in the prior art, for realizing 360 degree on a large scale, detection range far away, and static infrared heat source is also detected, effectively solve comfortableness in intelligent air condition with energy-conservation conflict and safety supervision system in the problem of high rate of false alarm, for this reason, the technical solution used in the present invention is, the feature extraction of human body thermal source and method of discrimination based on infrared thermal release electric information, comprise the steps: to adopt stepper motor to drive single pyroelectric infrared detector uniform rotation, realize remote, 360 degree on a large scale, the detection of stationary object; Investigative range is disk, and the sample of the human body collecting and non-human thermal source is carried out to wavelet packet analysis, by the signal feature of signal Wavelet Packet Entropy, utilizes BP neural network to carry out 5 folding cross validations, thereby completes the differentiation of human body and non-human thermal source.

Fresnel Lenses is installed in the place ahead at pyroelectric infrared detector, and two effects of Fresnel Lenses are: the one, and focussing force, is about to heat and releases infrared signal refraction, is reflected on pyroelectric infrared detector; Second effect is blind area and the visible range constantly alternately changing being divided in search coverage, the rpyroelectric infrared signal that the motive objects physical efficiency that makes to enter search coverage changes on pyroelectric infrared detector with the form of temperature variation, thereby output voltage signal.

Carry out wavelet packet analysis, by the signal feature concrete steps of signal Wavelet Packet Entropy, be: the globalize feature that obtains signal spectrum after processing by Fourier transform FFT, adopt Time-Frequency Analysis Method, time domain and frequency domain information are joined together, extract the time-frequency characteristics of signal; If

l ²(R) represent square-integrable real number space, i.e. the signal space of finite energy, t represents time variable, and R is real number, and its Fourier transform is ψ (ω), when ψ (ω) satisfies condition:

time, claim

it is a wavelet or for female small echo, by generating function

after flexible and translation, just can access a wavelet sequence, for continuous situation, the wavelet sequence obtaining is

Wherein, a is contraction-expansion factor, and b is shift factor, and for discrete situation, wavelet sequence is

For arbitrary function f (t) ∈ L ²(R), its continuous wavelet transform is

Be inversely transformed into

Adopt db1 small echo to carry out the analysis of signal;

A represents low frequency signal, and D represents high-frequency signal, and sequence number below of letter represents the number of plies of WAVELET PACKET DECOMPOSITION, scale parameter namely, and this decomposition has following relation:

S=AAA3+DAA3+ADA3+DDA3+AAD3+DAD3+ADD3+DDD3；

Adopt the non-normalized Shannon entropy of wavelet packet coefficient: establish c _ijthe coefficient of wavelet decomposition that represents j node of i layer, the non-normalized Shannon entropy of wavelet coefficient is:

$E\left(c_{\mathrm{ij}}\right)=\mathrm{\Σ}{c}_{\mathrm{ij}}^2\log \left(c_{\mathrm{ij}}^2\right)---\left(6\right)$

And agreement 0log0=0

By the Wavelet Packet Entropy composition characteristic vector of each node of j layer, carry out next step Classification and Identification; Adopt the Wavelet Packet Entropy composition characteristic vector of each node of the 3rd layer of db1 small echo, describe human body and non-human PIR signal characteristic, send in BP neural network and classify.

Utilize BP neural network to carry out 5 folding cross validations, thereby the differentiation that completes human body and non-human thermal source is specially, and adopts S type logic NOT linear function f (x)=1/ (1+e ^-x), its calculation procedure is as follows:

(1) initialization weights W and threshold value θ, be arranged to less random number all weights and threshold value;

(2) provide training sample x _i=(x ₀, x ₂... x _m-1), the object vector D of training sample is set _i=(d ₀, d ₁..., d _n-1), x wherein _irepresent i sample or output mode;

(3) by S type function and following computing formula, calculate the output x of each hidden layer _joutput valve y with output layer _k, the output of input node equals its input, supposes that input layer has m unit, and hidden layer has m ₁individual unit, output layer has n unit:

x _j＝f(∑w _ipx _i-θ _p) 0≤p≤m ₁-1 (7)

y _k＝f(∑w _pqx _q-θ _q) 0≤q≤n-1 (8)

(4) adjust weights.Use recursive algorithm to start reverse propagated error until the first hidden layer from output layer, and with following formula adjustment weights:

w _ip(t+1)＝w _ip(t)+ηδ _px _i (9)

In formula: w _ip(t)-at time t by hidden layer (or input layer) node i the weights to output layer (or hidden layer) node p;

X _ithe output of-node i; η δ _px _i-gain term; δ _pthe error term of-node p;

(5) ask system average error.To each pattern, to i, its error sum of squares is:

$E_i=\frac{1}{2}\left(\underset{q=0}{\overset{n-1}{\mathrm{\Σ}}}{(d_0-y_q)}^2\right)---\left(10\right)$

The average error of system is (supposing to have l sample): $E_i=\underset{i=0}{\overset{l}{\mathrm{\Σ}}}{E}_i=\frac{1}{2p}\left(\underset{i=0}{\overset{l-1}{\mathrm{\Σ}}}\underset{p=0}{\overset{n-1}{\mathrm{\Σ}}}{(d_{\mathrm{iq}}-y_{\mathrm{iq}})}^2\right)---\left(11\right)$

In formula: d _iq---the desired output of q output layer node of i input pattern; y _pq---corresponding calculating exported, if error does not meet the demands, repeats the calculating of (2)～(5) step, until the average error of system is less than the requirement of regulation;

Wherein, employing speed is adjusted, network structure regulation changes hidden layer neuron number, revise frequency of training and change neural network, the human body that adopts the method for five folding cross validations to solve different distance is differentiated discrimination: original sample collection is by random K the subset that be divided into, at K son, concentrate and select in turn one as checking collection, remaining K-1 subset is as training set, K subset all served as checking collection once after, it is for K time K folding that cross-validation process has just been repeated, afterwards the result of K checking is averaged and just can be obtained cross validation rate one time, adopt 5-folding cross validation method, the data set that the characteristic parameter extracting is formed carries out computing, then use correct identification (PCR) to carry out the recognition performance of evaluation model, the computing formula of correct recognition rata is as follows:

$\mathrm{PCR}=\frac{N_c}{N}\×100\%---\left(12\right)$

Wherein, Nc is the correct sample number of identification, and N is total test specimens given figure.

Technical characterstic of the present invention and effect:

The invention provides and a kind ofly utilize pyroelectric infrared detector to carry out human body to sentence method for distinguishing.To contributing to solve comfortableness and energy-conservation conflicting in Air-conditioner design, and the higher problem of rate of false alarm in safe examination system.Spatial sensitivity and the investigative range of pyroelectric sensor restrict mutually.Therefore the design proposes to utilize driven by motor to carry out the scanning within the scope of 360 degree with the pyroelectric detector of Fresnel Lenses, realize remote, on a large scale, high sensitivity, the object that also can detect static thermal source.Infrared signal to the human body in environment and non-human thermal source (computer screen) gathers, all there are 60 human samples and 40 non-human samples in different distance place, extract the Wavelet Packet Entropy of human body thermal source and non-human thermal source as feature, send into BP neural network and carry out human body and the non-human thermal source discrimination that five folding cross validations obtain different distance place, as can be seen from Figure 10 discrimination is higher, the discrimination at 1 meter of is lower is due to PIR detector, to have the inclination angle of 55 degree, and there is the blind area of about 1 meter centre.6 meters～7 meters cannot detect computer screen, but still can detect human body.The discrimination of single feature is higher, strong proof the feasibility of this scheme.

Accompanying drawing explanation

Fig. 1 is entire block diagram of the present invention.

Fig. 2 RE200B type Pyroelectronic Sensor figure.

Fig. 3 Fresnel Lenses device schematic diagram.(a) output signal when focussing force of Fresnel Lenses (b) human body is by PIR.

Fig. 4 Fresnel Lenses blocks comparison diagram.(a) Fresnel Lenses of Fresnel Lenses (b) after blocking.

Fig. 5 PIR data acquisition software front panel

Fig. 6 experiment scene simulation drawing.

Fig. 7 scans the result in time domain figure after 360 degree.

Fig. 8 WAVELET PACKET DECOMPOSITION tree schematic diagram.

Fig. 9 neural network basic model.

Figure 10 BP neural network recognization rate.

Embodiment

The present invention can utilize single pyroelectric detector to realize 360 degree on a large scale, detection range far away, and static infrared heat source is also detected, the Wavelet Packet Entropy of extracting PIR signal is feature, sends into the differentiation that BP neural network is carried out human body.Thereby effectively solve comfortableness in intelligent air condition with energy-conservation conflict and safety supervision system in the problem of high rate of false alarm.

Because the thermal source of different shapes is when by PIR sensing range, will produce different signals, this is just for utilizing pyroelectric sensor to carry out human body and non-human detection provides theoretic foundation.Accordingly, adopt stepper motor to drive the single PIR detector uniform rotation with Fresnel Lenses, realize remote, 360 degree on a large scale, the detection of stationary object.Investigative range is the disk of 7.4 meters of radiuses.Surveying initial point place, to have radius be the blind area of 0.9 meter.The sample of the human body collecting and non-human thermal source is carried out to wavelet packet analysis, by the signal feature of signal Wavelet Packet Entropy, utilize BP neural network to carry out 5 folding cross validations, thereby complete the differentiation of human body and non-human thermal source, realize the differentiation of human body.The design's structured flowchart is as follows:

The ultimate principle of 1 pyroelectric infrared sensor

Human body is good infrared radiation source, and typical human body temperature is 37 ℃ or 98 ℉, and the radiation wavelength that can calculate human body through Wien's displacement law is that the sensitive wave length scope of 9～10 μ m. pyroelectric infrared sensors is conventionally in 8～12 μ m left and right.Pyroelectric effect refers to that the variation of temperature makes the atom site generation minor alteration of crystal structure inside, thereby causes the degree of polarization of material to change, and the change of this degree of polarization can make crystal two ends produce a voltage signal.A kind of electrothermic type infrared detector cell that pyroelectric infrared sensor utilizes pyroelectric effect principle to make just.The essence of pir sensor detection principle is to detect the variation that is radiated at the infrared light intensity on chip, and the temperature that causes pyroelectric crystal in sensor changes, and changes the faster pyroelectricity signal producing also larger.Conventionally by 2 opposite polarity sensing elements, be connected in series, and fit together with 1 high resistant and 1 field effect transistor.Because 2 sensing element polarity are contrary, connection signal will be cancelled out each other and not output.Therefore pyroelectric sensor is to the variation of environment temperature, background radiation, and the random noise of self temperature variation and vibrated generation all has good compensating action, only to the movement of human body or motion sensitive, makes sensor reliable and stable in actual use simultaneously.The design adopts RE200B type Pyroelectronic Sensor, in kind as Fig. 2.

2 human body pyroelectricity infrared information acquisition systems

In order to improve detector sensitivity and detection range, Fresnel Lenses of the place ahead of the common detector installing or diffraction optics type focus lamp etc.The place ahead at pyroelectric sensor in the design is provided with Fresnel Lenses, and its two effects are: the one, and focussing force, is about to heat and releases infrared signal refraction (reflection) on pir sensor, and this can make the detection range of sensor greatly increase as shown in Figure 3; Second effect is blind area (dark space) and visible range (area pellucida) constantly alternately changing being divided in search coverage, the rpyroelectric infrared signal that the motive objects physical efficiency that makes to enter search coverage changes on pir sensor with the form of temperature variation, thereby output voltage signal.

This research adopts 8120 type Fresnel Lenses, and this lens volume is little, is easy to installation and investigation depth darker.During experiment, carry out the requirement that partial occlusion meets approximate rectangle surveyed area.The rectangular roomy width (50cm) that is about a people.Because the signal of pyroelectric sensor output is very faint, and signal also comprises other and disturbs, therefore need to be to the amplifying and filtering processing containing noisy ultra-weak electronic signal of sensor output, to meet the needs of signal processing.

Digital-to-analogue (A/D) conversion equipment, i.e. data collecting card, the NI USB-6251 data collecting card that adopts American National instrument company (National Instruments, NI) to produce.The software platform of system data collection is set up based on LabVIEW, a kind of Software Development Platform for Virtual Instruments of Ta Shi America NI company development.The store path, sampling rate and the hits that on the interface of native system design, specifically comprise data, and the demonstration of the real-time waveform of signal, the overall waveform of time-domain signal shows and the parts such as frequency spectrum demonstration of signal after Fourier transform.

3 rpyroelectric infrared data acquisition and signal characteristic abstractions

The design utilizes the static human body in different radii place and non-human thermal source (computer screen) is carried out to the collection of PIR signal, utilize afterwards wavelet packet analysis to come the Wavelet Packet Entropy of picked up signal as feature, utilize afterwards BP(Back Propagation) neural network carries out the differentiation of human body as sorter.

3.1 rpyroelectric infrared information acquisitions

Pyroelectric infrared sensor also has the existence of many problems, and for example investigative range is little, can not survey static object etc.In order to solve two problems above-mentioned, and in order to improve detection sensitivity, propose with the uniform rotation of driven by motor pyroelectric infrared sensor.The design selects 42H2P3412A4 type two-phase stepping motor, drive single PIR detector to realize the detection within the scope of 360 degree, run-down needs 30s, wherein to be placed on height be on the support of 2.7 meters to PIR, the elevation angle is 55 degree, and carry out partial occlusion, realized radius and be the disk scanning of 7.4 meters.Computer screen is non-human thermal source.It is 1 to 7 meter of that human body and non-human thermal source all stand at random from the center of circle, every 0.5 meter of collection of all carrying out single pass and data.Fig. 6 is experiment scene simulation drawing, time domain waveform when Fig. 7 is driven by motor pyroelectric detector run-down.Now to be randomly dispersed in radius be on the circular arc of 2 meters to the non-human thermal source of human body thermal source and target.Wherein the 3rd maximum point is the less infrared heat source that insert row and computer charger produce, and the 4th maximum point is the computer that image data is used, and it is 1.5 meters apart from initial point.So amplitude is large compared with first and second maximum point (being computer screen).Five, six, seven maximum point behaviour body heat sources.Only by front two computers screen, feature extraction and differentiation after rear three personal accomplishment samples carry out.

3.2 PIR signal characteristic abstractions based on Wavelet Packet Entropy

After processing by FFT, can obtain the globalize feature of signal spectrum, but in spectrogram, not comprise temporal information.And the human body signal obtaining in experiment belongs to non-stationary signal, therefore, adopted Time-Frequency Analysis Method, time domain and frequency domain information are joined together, extract the time-frequency characteristics of signal.Wavelet analysis is widely used Time-Frequency Analysis Method, can obtain good temporal resolution and frequency resolution simultaneously.If

l ²(R) represent square-integrable real number space (being the signal space of finite energy), t represents time variable, and R is real number, and its Fourier transform is ψ (ω).When ψ (ω) satisfies condition:

time, we claim

it is a wavelet or for female small echo.By generating function

after flexible and translation, just can access a wavelet sequence.For continuous situation, the wavelet sequence obtaining is

Wherein, a is contraction-expansion factor, and b is shift factor.For discrete situation, wavelet sequence is

For arbitrary function f (t) ∈ L ²(R), its continuous wavelet transform is

Be inversely transformed into

Wherein

by expression formula (1), obtained W _f(a, b) is continuous wavelet transform, and a is contraction-expansion factor, and b is shift factor,

In practical application, wavelet basis function

(t) do not have unique selection, time-domain analysis requires wavelet function to have lower frequency resolution at HFS, and frequency-domain analysis requires same wavelet function to have higher frequency resolution in low frequency part.Because raw data or function can launch with a series of wavelet coefficients, so wavelet coefficient can be used as and analyze and a kind of means of processing signals, selects best wavelet function, contributes to better to describe the time-frequency characteristics of signal.In the design, select db1 small echo.Because db1 small echo is Harr small echo, it is that rigidity is symmetrical and antisymmetric, and when infrared heat source moves in PIR surveyed area, the polarity of waveform changes from forward is negative, and output signal, in the positive and negative alternately appearance of time domain, has good antisymmetry feature.This symmetrical feature of different thermals source is different.Therefore the analysis that, adopts db1 small echo to carry out signal can obtain good result

Wavelet packet analysis (WPA) is multiresolution analysis and the Mallat algorithm based on classical, adopts orthogonal wavelet basis function to divide layer by layer signal.Wavelet packet analysis is compared with wavelet transform, not only low frequency signal is decomposed, and the detail signal of high frequency has also passed through multilayer division.For the WAVELET PACKET DECOMPOSITION of a N layer, can produce 2 ⁿindividual different wavelet coefficient sequence (or node), the wavelet transform of the same number of plies only has (3N+1) individual.Yet after this down-sampling process,, can't there is redundancy in whole wavelet coefficient or identical.Its WAVELET PACKET DECOMPOSITION tree as shown in Figure 8.

In Fig. 8, A represents low frequency signal, and D represents high-frequency signal, and sequence number below of letter represents the number of plies of WAVELET PACKET DECOMPOSITION, scale parameter namely, and this decomposition has following relation:

S=AAA3+DAA3+ADA3+DDA3+AAD3+DAD3+ADD3+DDD3。

Entropy originates from information theory, can describe complexity and the random degree of random time sequence.The information of the potential dynamic evolution process of signal can be provided.Wavelet package transforms and entropy can effectively be obtained to the information of describing signal characteristic in conjunction with the Wavelet Packet Entropy obtaining.Entropy has various definitions form, adopts the non-normalized Shannon entropy of wavelet packet coefficient herein.If c _ijthe coefficient of wavelet decomposition that represents j node of i layer, the non-normalized Shannon entropy of wavelet coefficient is:

$E\left(c_{\mathrm{ij}}\right)=\mathrm{\Σ}{c}_{\mathrm{ij}}^2\log \left(c_{\mathrm{ij}}^2\right)---\left(6\right)$

And agreement 0log0=0

By the Wavelet Packet Entropy composition characteristic vector of each node of j layer, carry out next step Classification and Identification.Adopt the Wavelet Packet Entropy composition characteristic vector of each node of the 3rd layer of db1 small echo herein, describe human body and non-human PIR signal characteristic, send in BP neural network and classify.

4 human body and non-human differentiations based on BP neural network

Because artificial neural network comes from the simulation to cranial nerve, there are the very strong ability that is adapted to complex environment and multi objective control to have good self-learning capability, and the characteristic of can arbitrary accuracy approaching any non-linear continuous function.Make prophesy, at the beginning of 21 century, artificial neural network's theory will have larger development, and the large step that promotes science and technology is advanced in its application.BP(Back Propagation) neural network is current most widely used the most ripe general a kind of neural network model, as shown in Figure 9, it is constructed by hierarchical structure, it comprises an input layer, an output layer and one or more hidden layer, and the node in one deck is only connected with each node of lower one deck of this layer of next-door neighbour.Wherein every one deck all comprises several neurons, between the neuron in same layer, does not interknit, and the number of plies of each layer of neuron number and hidden layer should be depending on particular problem.

BP neural network is the neural network model of a supervised learning (training).BP neural network comprises two processes, training process and test processs.Training process is that the training sample of some is provided to network, the actual output of network and desired output vector are compared, by revising the weight coefficient between each layer of neuron, make the error between desired output and actual output reach minimum, determine weights coefficient, whole network has also been determined thereupon.Test process is sent test sample book into neural network exactly, and the output vector of trying to achieve is test result.

If its transfer function must be continuously differentiable nonlinear function, generally adopt S type logic NOT linear function f (x)=1/ (1+e ^-x).Its calculation procedure is as follows:

(1) initialization weights W and threshold value θ, be arranged to less random number all weights and threshold value.

(2) provide training sample x _i=(x ₀, x ₂... x _m-1), the object vector D of training sample is set _i=(d ₀, d ₁..., d _n-1), wherein subscript i represents i sample or output mode.

(3) by S type function and following computing formula, calculate the output x of each hidden layer _joutput valve y with output layer _k, the output of input node equals its input, supposes that input layer has m unit, and hidden layer has m ₁individual unit, output layer has n unit:

x _j＝f(∑w _ipx _i-θ _p) 0≤p≤m ₁-1 (7)

y _q＝f(∑w _pqx _q-θ _q) 0≤q≤n-1 (8)

(4) adjust weights.Use recursive algorithm to start reverse propagated error until the first hidden layer from output layer, and with following formula adjustment weights:

w _ip(t+1)＝w _ip(t)+ηδ _px _i (9)

In formula: w _ip(t)-at time t by hidden layer (or input layer) node i the weights to output layer (or hidden layer) node p;

X _ithe output of-node i; η δ _qx _i-gain term; δ _qthe error term of-node q;

(5) ask system average error.To each pattern, to i, its error sum of squares is:

$E_i=\frac{1}{2}\left(\underset{q=0}{\overset{n-1}{\mathrm{\Σ}}}{(d_q-y_q)}^2\right)---\left(10\right)$

The average error of system is (supposing to have l sample):

$E_i=\underset{i=0}{\overset{l}{\mathrm{\Σ}}}{E}_i=\frac{1}{2p}\left(\underset{i=0}{\overset{l-1}{\mathrm{\Σ}}}\underset{p=0}{\overset{n-1}{\mathrm{\Σ}}}{(d_{\mathrm{iq}}-y_{\mathrm{iq}})}^2\right)---\left(11\right)$

In formula: d _iq---the desired output of q output layer node of i input pattern; y _pq---corresponding calculating exported.If error does not meet the demands, repeat the calculating of (2)～(5) step, until the average error of system is less than the requirement of regulation.

Under study for action.We adopt speed adjustment, network structure regulation (changing hidden layer neuron number), revise the change neural networks such as frequency of training, because the sample of human body thermal source and non-human thermal source is less, the human body that we have adopted the method for five folding cross validations to solve different distance is differentiated discrimination.Cross validation is also known as cycle criterion sometimes, is a kind of a kind of method of independent data sets that how to be applied to for assessing statistic analysis result.Cross validation is mainly used in the situation of sample prediction, and need to evaluate in actual applications the accuracy of a forecast model.K folding cross validation (K-fold Cross-validation) is a common type in cross validation, and in K folding cross validation, original sample collection is by random K the subset that be divided into.At K son, concentrate and select in turn one as checking collection, a remaining K-1 subset is as training set.K subset all served as checking collection once after, cross-validation process has just been repeated (being K folding) K time, afterwards the result of verifying for K time is averaged and just can be obtained cross validation rate one time.The feature of this method is that each random subset producing was selected as training set and test set, and had all obtained the accuracy of a checking at every turn.Adopt 5-folding cross validation method herein, the data set that the characteristic parameter extracting is formed carries out computing, then uses correct identification (PCR) to carry out the recognition performance of evaluation model.The computing formula of correct recognition rata is as follows:

$\mathrm{PCR}=\frac{N_c}{N}\×100\%---\left(12\right)$

Wherein, Nc is the correct sample number of identification, and N is total test specimens given figure.

The design provides a kind of and has utilized pyroelectric infrared detector to carry out human body to sentence method for distinguishing.To contributing to solve comfortableness and energy-conservation conflicting in Air-conditioner design, and the higher problem of rate of false alarm in safe examination system.Spatial sensitivity and the investigative range of pyroelectric sensor restrict mutually.Therefore the design proposes to utilize driven by motor to carry out the scanning within the scope of 360 degree with the pyroelectric detector of Fresnel Lenses, realize remote, on a large scale, high sensitivity, the object that also can detect static thermal source.Infrared signal to the human body in environment and non-human thermal source (computer screen) gathers, all there are 60 human samples and 40 non-human samples in different distance place, extract the Wavelet Packet Entropy of human body thermal source and non-human thermal source as feature, send into BP neural network and carry out human body and the non-human thermal source discrimination that five folding cross validations obtain different distance place, as can be seen from Figure 10 discrimination is higher, the discrimination at 1 meter of is lower is due to PIR detector, to have the inclination angle of 55 degree, and there is the blind area of about 1 meter centre.6 meters～7 meters cannot detect computer screen, but still can detect human body.The discrimination of single feature is higher, strong proof the feasibility of this scheme.

It is a complexity and challenging research topic that human body is differentiated, and relates to the knowledge of a plurality of research subjects, as computer vision, image processing, pattern-recognition etc.The design has proposed a kind of new human body and has differentiated scheme, is intended to be applied in Smart Home, solves conflicting of comfortableness and energy-saving index in air-conditioning.Also for human detection rate of false alarm in safety supervision system is higher, provide a solution, and obtain considerable social benefit simultaneously.Optimum implementation intends adopting technological cooperation or product development.

Claims (4)

1. the feature extraction of human body thermal source and the method for discrimination based on infrared thermal release electric information, it is characterized in that, comprise the steps: to adopt stepper motor to drive single pyroelectric infrared detector uniform rotation, realize remote, 360 degree on a large scale, the detection of stationary object; Investigative range is disk, and the sample of the human body collecting and non-human thermal source is carried out to wavelet packet analysis, by the signal feature of signal Wavelet Packet Entropy, utilizes BP neural network to carry out 5 folding cross validations, thereby completes the differentiation of human body and non-human thermal source.

2. the feature extraction of human body thermal source and the method for discrimination based on infrared thermal release electric information as claimed in claim 1, it is characterized in that, Fresnel Lenses is installed in the place ahead at pyroelectric infrared detector, two effects of Fresnel Lenses are: the one, and focussing force, is about to heat and releases infrared signal refraction, is reflected on pyroelectric infrared detector; Second effect is blind area and the visible range constantly alternately changing being divided in search coverage, the rpyroelectric infrared signal that the motive objects physical efficiency that makes to enter search coverage changes on pyroelectric infrared detector with the form of temperature variation, thereby output voltage signal.

3. the feature extraction of human body thermal source and the method for discrimination based on infrared thermal release electric information as claimed in claim 1, it is characterized in that, carry out wavelet packet analysis, by the signal feature concrete steps of signal Wavelet Packet Entropy, be: the globalize feature that obtains signal spectrum after processing by Fourier transform FFT, adopt Time-Frequency Analysis Method, time domain and frequency domain information are joined together, extract the time-frequency characteristics of signal; If

l ²(R) represent square-integrable real number space, i.e. the signal space of finite energy, t represents time variable, and R is real number, and its Fourier transform is ψ (ω), when ψ (ω) satisfies condition:

time, claim

it is a wavelet or for female small echo, by generating function after flexible and translation, just can access a wavelet sequence, for continuous situation, the wavelet sequence obtaining is

Wherein, a is contraction-expansion factor, and b is shift factor, and for discrete situation, wavelet sequence is

For arbitrary function f (t) ∈ L ²(R), its continuous wavelet transform is

Be inversely transformed into

Adopt db1 small echo to carry out the analysis of signal;

A represents low frequency signal, and D represents high-frequency signal, and sequence number below of letter represents the number of plies of WAVELET PACKET DECOMPOSITION, scale parameter namely, and this decomposition has following relation:

S=AAA3+DAA3+ADA3+DDA3+AAD3+DAD3+ADD3+DDD3；

Adopt the non-normalized Shannon entropy of wavelet packet coefficient: establish c _ijthe coefficient of wavelet decomposition that represents j node of i layer, the non-normalized Shannon entropy of wavelet coefficient is:

$E\left(c_{\mathrm{ij}}\right)=\mathrm{\Σ}{c}_{\mathrm{ij}}^2\log \left(c_{\mathrm{ij}}^2\right)---\left(6\right)$

And agreement 0log0=0

By the Wavelet Packet Entropy composition characteristic vector of each node of j layer, carry out next step Classification and Identification; Adopt the Wavelet Packet Entropy composition characteristic vector of each node of the 3rd layer of db1 small echo, describe human body and non-human PIR signal characteristic, send in BP neural network and classify.

4. the feature extraction of human body thermal source and the method for discrimination based on infrared thermal release electric information as claimed in claim 1, it is characterized in that, utilize BP neural network to carry out 5 folding cross validations, thereby the differentiation that completes human body and non-human thermal source is specially, adopt S type logic NOT linear function f (x)=1/ (1+e ^-x), its calculation procedure is as follows:

(1) initialization weights W and threshold value θ, be arranged to less random number all weights and threshold value;

(2) provide training sample x _i=(x ₀, x ₂... x _m-1), the object vector D of training sample is set _i=(d ₀, d ₁..., d _n-1), x wherein _irepresent i sample or output mode;

(3) by S type function and following computing formula, calculate the output x of each hidden layer _joutput valve y with output layer _k, the output of input node equals its input, supposes that input layer has m unit, and hidden layer has m ₁individual unit, output layer has n unit:

x _j＝f(∑w _ipx _i-θ _p) 0≤p≤m ₁-1 (7)

y _k＝f(∑w _pqx _q-θ _q) 0≤q≤n-1 (8)

(4) adjust weights, use recursive algorithm to start reverse propagated error until the first hidden layer from output layer, and with following formula adjustment weights:

w _ip(t+1)＝w _ip(t)+ηδ _px _i (9)

In formula: w _ip(t)-at time t by hidden layer (or input layer) node i the weights to output layer or hidden layer node p;

X _ithe output of-node i; η δ _px _i-gain term; δ _pthe error term of-node p;

(5) ask system average error.To each pattern, to i, its error sum of squares is:

$E_i=\frac{1}{2}\left(\underset{q=0}{\overset{n-1}{\mathrm{\Σ}}}{(d_q-y_q)}^2\right)---\left(10\right)$

The average error of system is to suppose to have l sample: $E_i=\underset{i=0}{\overset{l}{\mathrm{\Σ}}}{E}_i=\frac{1}{2p}\left(\underset{i=0}{\overset{l-1}{\mathrm{\Σ}}}\underset{p=0}{\overset{n-1}{\mathrm{\Σ}}}{(d_{\mathrm{iq}}-y_{\mathrm{iq}})}^2\right)---\left(11\right)$

In formula: d _iq---the desired output of q output layer node of i input pattern; y _pq---corresponding calculating exported, if error does not meet the demands, repeats the calculating of (2)～(5) step, until the average error of system is less than the requirement of regulation;

Wherein, employing speed is adjusted, network structure regulation changes hidden layer neuron number, revise frequency of training and change neural network, the human body that adopts the method for five folding cross validations to solve different distance is differentiated discrimination: original sample collection is by random K the subset that be divided into, at K son, concentrate and select in turn one as checking collection, remaining K-1 subset is as training set, K subset all served as checking collection once after, it is for K time K folding that cross-validation process has just been repeated, afterwards the result of K checking is averaged and just can be obtained cross validation rate one time, adopt 5-folding cross validation method, the data set that the characteristic parameter extracting is formed carries out computing, then use correct identification (PCR) to carry out the recognition performance of evaluation model, the computing formula of correct recognition rata is as follows:

$\mathrm{PCR}=\frac{N_c}{N}\×100\%---\left(12\right)$

Wherein, Nc is the correct sample number of identification, and N is total test specimens given figure.

Images