CN106874961A - A kind of indoor scene recognition methods using the very fast learning machine based on local receptor field - Google Patents

Abstract

Use the very fast learning machine (LRF ELM) based on local receptor field come the method for processing indoor scene classification problem the present invention relates to a kind of, belong to the indoor scene recognition methods of mobile robot.The method includes:1) gather as 2 D radar informations of training sample；2) gather as 2 D radar informations of test sample；3) binary image information extraction is carried out to radar information；4) training LRF ELM networks obtain optimal output weight；5) verify that the method carries out the accuracy rate of scene Recognition by test set.On the basis of neutral net, the correctness to the indoor scene identification based on 2 D radar informations judges the present invention, shortens operation time, substantially increases the efficiency of scene Recognition.

Description

A kind of indoor scene recognition methods using the very fast learning machine based on local receptor field

Technical field

Indoor scene point is processed using the very fast learning machine (LRF-ELM) based on local receptor field the present invention relates to a kind of The method of class problem, belongs to the indoor scene recognition methods field of mobile robot.

Background technology

Indoor scene identification produces far-reaching influence as current study hotspot to daily life.It should Video monitoring, smart home, the daily production operation of robot and rescue of hazardous environment etc. are essentially consisted in value.

With modern sensor, control and the development of artificial intelligence technology, scientific research personnel is to based on computer vision and base Extensive research is carried out in the scene recognition method of the broad aspect of depth perception two." one kind is based on the patent of invention of Zhejiang University The depth map of the present frame that the indoor scene localization method of hybrid camera " is shot with hybrid camera and cromogram and training Good recurrence forest, calculates the corresponding world coordinates of current camera, completes indoor scene positioning, but its effect receives light shadow Sound is larger.

The content of the invention

The purpose of the present invention is the weak point for overcoming conventional art, proposes a kind of fast and effectively indoor scene positioning side Method, realizes that the indoor scene based on 2-D radar informations is positioned on the basis of the very fast learning machine based on local receptor field, improves The efficiency and accuracy rate of indoor scene identification.

A kind of indoor scene localization method using the very fast learning machine based on local receptor field proposed by the present invention, including Following steps:

(1) collection if the number of training sample is N, then obtains training sample as the radar information of the scene of training sample Notebook data collection S_trExpression formula be：

S_tr={ S_tr1,S_tr2,…,S_trN}

Wherein S_tr1,S_tr2,…,S_trNTraining sample data collection S is represented respectively_trIn first training sample, second ultrasound Training sample ... n-th training sample.In different scenes, the training sample number for being gathered is roughly the same；

(2) collection of radar information is carried out to the test sample scene that needs are identified.If the number of test sample is M, then obtain ultrasonic tesint sample data set S_teExpression formula be：

S_te={ S_te1,S_te2,…,S_teM}

Wherein S_te1,S_te2,…,S_teMTest sample data set S is represented respectively_teIn first test sample, second test Sample ... m-th test sample.In different scenes, the training sample number for being gathered is roughly the same.M and N are respectively training sample The number of this number and test sample, generally M≤N；

(3) to radar range finding training set S_trSample information carry out feature extraction, concrete processing procedure is as follows：

(3-1) note training sample set S_trIn any one training sample be S_I, 1≤I≤N, S_IIt is one to be swept by radar Retouch the one-dimensional characteristic vector that the radar data for obtaining for a week is constituted, i.e. S_I=[S_I.1, S_I.2..., S_I.l], wherein S_I.1, S_I.2..., S_I.lThe l radar data of sampled point in single pass is represented, it is converted into polar coordinate image by this group of radar data；

(3-2) extracts polar coordinate image in rectangular coordinate system, and using the center of circle of polar coordinate image as rectangular image Center, point red filling is carried out to contoured interior；

(3-3) carries out binary conversion treatment to the rectangle red blank map obtained in (3-2) with white as background, make its into It is black white image, obtains view data, obtains the sample S of new training set_I', complete the feature extraction to training set sample；

And then obtain new training sample set S_tr'：

S_tr'={ S_tr1',S_tr2',…,S_trk',…,S_trN'}；

Wherein, S_tr1',S_tr2',…,S_trk',…,S_trN' represent respectively through asking the test obtained after binary image data Training set S_tr' in first training sample, second training sample ..., k-th training sample ..., n-th training sample, N is number of training；

(3-4) gives training set S_tr' in sample from different type room set different labels, and constitute label matrix T；

(4) to radar range finding test set S_teSample information carry out feature extraction, concrete processing procedure is as follows：

(4-1) note test sample collection S_teIn any one training sample be S_J, 1≤J≤M, S_JIt is one to be swept by radar Retouch the one-dimensional characteristic vector that the radar data for obtaining for a week is constituted, i.e. S_J=[S_J.1, S_J.2..., S_J.l], wherein S_J.1, S_J.2..., S_J.lThe l radar data of sampled point in single pass is represented, it is converted into polar coordinate image by this group of radar data；

(4-2) extracts polar coordinate image in rectangular coordinate system, and using the center of circle of polar coordinate image as rectangular image Center, point red filling is carried out to contoured interior；

(4-3) carries out binary conversion treatment to the rectangle red blank map obtained in (4-2) with white as background, make its into It is black white image, obtains view data, obtains the sample S of new training set_J', complete the feature extraction to training set sample；

And then obtain new test sample collection S_te'：

S_te'={ S_te1',S_te2',…,S_tek',…,S_teM'}；

Wherein, S_te1',S_te2',…,S_tek',…,S_teM' represent respectively through asking the test obtained after binary image data Collection S_te' in first test sample, second test sample ..., k-th test sample ..., m-th test sample, M is Test sample number；

(4-4) gives test set S in the way of in (3-4)_te' in sample from different type room set different marks Sign, and constitute label matrix T'；

(5) by training set S_tr' and corresponding label matrix T, test set S_te' and corresponding label matrix T' as being based on The input of the very fast learning machine of local receptor field, and the relevant parameters such as convolutional layer, pond layer are set；

(5-1) generates input weight W at random_i=[w_i1,,,w_in]^TWith the biasing b of Hidden unit_i=[b_i1,…b_in]^T, it is right Initial weight is orthogonalized, and obtains new input weightIf training set sample input size is (d × d), receptive field size It is (r × r), k-th value c of convolution node (i, j) of characteristic pattern_i,j,kCalculated by following formula：

(5-2) carries out square root pond to characteristic pattern, and pond size e represents Chi Hua centers to the distance on side, h_p,q,kRepresent Combined joint (p, q) in k-th pond figure, is calculated as follows：

The value that (5-3) simply connects all combined joints forms a row vector, and N number of training set input sample Row vector is put together, obtains combination layer matrix H, and output weight beta is calculated by following formula：

As N ＞ K (d-r+1)²,

As N≤K (d-r+1)²,

(5-4) input weightIt is constant, to test set S_te' sample carry out and identical convolution in (5-1) and (5-2) With pond process, combination layer H' is obtained, imputation method is Y to the Tag Estimation of test set sample, be can be calculated：

Y=H'* β

Prediction label Y is contrasted with the given label T' of test set, the accuracy of scene Recognition is drawn.

It is proposed by the present invention based on 2-D radar informations, using the indoor scene of the very fast learning machine based on local receptor field Recognition methods, substantially reduces operation time, improves the efficiency of indoor scene identification.And the inventive method is simple and reliable, With very strong practicality.

Brief description of the drawings

Fig. 1 is algorithm flow chart used by the present invention.

Fig. 2 is the schematic diagram of very fast learning machine (LRF-ELM) of the algorithm used based on local receptor field in the present invention.

Fig. 3 is the procedure chart of binary image information extraction in carrying out characteristic extraction procedure to training set and test set.

Specific embodiment

It is proposed by the present invention a kind of based on 2-D radar informations, using the interior of the very fast learning machine based on local receptor field Scene recognition method, specific embodiment further describes as follows.

(1) radar sensor is installed in mobile robot, the radar information as the scene of training sample is gathered, if instruction The number for practicing sample is N, then obtain training sample data collection S_trExpression formula be：

S_tr={ S_tr1,S_tr2,…,S_trN}

Wherein S_tr1,S_tr2,…,S_trNTraining sample data collection S is represented respectively_trIn first training sample, second ultrasound Training sample ... n-th training sample.In different scenes, the training sample number for being gathered is roughly the same.

(2) collection of radar information is carried out to the test sample scene that needs are identified.If the number of test sample is M, then obtain ultrasonic tesint sample data set S_teExpression formula be：

S_te={ S_te1,S_te2,…,S_teM}

Wherein S_te1,S_te2,…,S_teMTest sample data set S is represented respectively_teIn first test sample, second test Sample ... m-th test sample.In different scenes, the training sample number for being gathered is roughly the same.M and N are respectively training sample The number of this number and test sample, generally M≤N.

(3) feature extraction is carried out to radar range finding sample information, concrete processing procedure is as follows：

(3-1) note training sample set S_trIn any one training sample be S_I, 1≤I≤N, S_IIt is one to be swept by radar Retouch the one-dimensional characteristic vector that the radar data for obtaining for a week is constituted, i.e. S_I=[S_I.1, S_I.2..., S_I.l], wherein S_I.1, S_I.2..., S_I.lThe l radar data of sampled point in single pass is represented, it is converted into polar coordinate image by this group of radar data；

(3-2) extracts polar coordinate image in rectangular coordinate system, and using the center of circle of polar coordinate image as rectangular image Center, point red filling is carried out to contoured interior；

(3-3) carries out binary conversion treatment to the rectangle red blank map obtained in (3-2) with white as background, make its into It is black white image, obtains view data, obtains the sample S of new training set_I', complete the feature extraction to training set sample.

And then obtain new training sample set S_tr'：

S_tr'={ S_tr1',S_tr2',…,S_trk',…,S_trN'}

Wherein, S_tr1',S_tr2',…,S_trk',…,S_trN' represent respectively through asking the test obtained after binary image data Training set S_tr' in first training sample, second training sample ..., k-th training sample ..., n-th training sample, N is number of training；

(3-4) gives training set S_tr' in sample from different type room set different labels, such as corridor is 1, bathroom It is 2, bedroom is 3 etc., and constitutes label matrix T.

(4-1) note test sample collection S_teIn any one training sample be S_J, 1≤J≤M, S_JIt is one to be swept by radar Retouch the one-dimensional characteristic vector that the radar data for obtaining for a week is constituted, i.e. S_J=[S_J.1, S_J.2..., S_J.l], wherein S_J.1, S_J.2..., S_J.lThe l radar data of sampled point in single pass is represented, it is converted into polar coordinate image by this group of radar data；

(4-2) extracts polar coordinate image in rectangular coordinate system, and using the center of circle of polar coordinate image as rectangular image Center, point red filling is carried out to contoured interior；

(4-3) carries out binary conversion treatment to the rectangle red blank map obtained in (4-2) with white as background, make its into It is black white image, obtains view data, obtains the sample S of new training set_J', complete the feature extraction to training set sample.

And then obtain new test sample collection S_te'：

S_te'={ S_te1',S_te2',…,S_tek',…,S_teM'}

Wherein, S_te1',S_te2',…,S_tek',…,S_teM' represent respectively through asking the test obtained after binary image data Collection S_te' in first test sample, second test sample ..., k-th test sample ..., m-th test sample, M is Test sample number；

(4-4) gives test set S in the way of in (3-4)_te' in sample from different type room set different marks Sign, and constitute label matrix T'.

(5) by training set S_tr' and corresponding label matrix T, test set S_te' and corresponding label matrix T' as being based on The input of the very fast learning machine of local receptor field, and the relevant parameters such as convolutional layer, pond layer are set.

(5-1) generates input weight W at random_i=[w_i1,,,w_in]^TWith the biasing b of Hidden unit_i=[b_i1,…b_in]^T, it is right Initial weight is orthogonalized, and obtains new input weightIf training set sample input size is (d × d), receptive field size It is (r × r), k-th value c of convolution node (i, j) of characteristic pattern_i,j,kCalculated by following formula：

(5-2) carries out square root pond to characteristic pattern, and pond size e represents Chi Hua centers to the distance on side, h_p,q,kRepresent Combined joint (p, q) in k-th pond figure, is calculated as follows：

The value that (5-3) simply connects all combined joints forms a row vector, and N number of training set input sample Row vector is put together, obtains combination layer matrix H, and output weight beta is calculated by following formula：

As N ＞ K (d-r+1)²,

As N≤K (d-r+1)²,

(5-4) input weightIt is constant, to test set S_te' sample carry out with identical convolution in (5-1) and (5-2) with Pond process, obtains combination layer H', and imputation method is Y to the Tag Estimation of test set sample, be can be calculated：

Y=H'* β

Finally, prediction label Y is contrasted with the given label T' of test set, is drawn the correct of this scene Recognition Rate.

Claims (1)

1. it is a kind of to realize that indoor scene knows method for distinguishing using the very fast learning machine (LRF-ELM) based on local receptor field, its It is characterised by, the method is comprised the following steps：

(1) collection if the number of training sample is N, then obtains number of training as the radar information of the scene of training sample According to collection S_trExpression formula be：

S_tr={ S_tr1,S_tr2,…,S_trN}

Wherein S_tr1,S_tr2,…,S_trNTraining sample data collection S is represented respectively_trIn first training sample, second ultrasound training Sample ... n-th training sample.In different scenes, the training sample number for being gathered is roughly the same；

(2) collection of radar information is carried out to the test sample scene that needs are identified.If the number of test sample is M, then Obtain ultrasonic tesint sample data set S_teExpression formula be：

S_te={ S_te1,S_te2,…,S_teM}

Wherein S_te1,S_te2,…,S_teMTest sample data set S is represented respectively_teIn first test sample, second test specimens This ... m-th test sample.In different scenes, the training sample number for being gathered is roughly the same.M and N are respectively training sample Number and test sample number, generally M≤N；

(3) to radar range finding training set S_trSample information carry out feature extraction, concrete processing procedure is as follows：

(3-1) note training sample set S_trIn any one training sample be S_I, 1≤I≤N, S_IIt it is one by radar scanning one week The one-dimensional characteristic vector that the radar data of acquisition is constituted, i.e. S_I=[S_I.1, S_I.2..., S_I.l], wherein S_I.1, S_I.2..., S_I.l The l radar data of sampled point in single pass is represented, it is converted into polar coordinate image by this group of radar data；

(3-2) extracts polar coordinate image in rectangular coordinate system, and using the center of circle of polar coordinate image as in rectangular image The heart, red filling is carried out to contoured interior point；

(3-3) carries out binary conversion treatment to the red blank map of the rectangle with white as background obtained in (3-2), becomes black White image, obtains view data, obtains the sample S of new training set_I', complete the feature extraction to training set sample；

And then obtain new training sample set S_tr'：

S_tr'={ S_tr1',S_tr2',…,S_trk',…,S_trN'}；

Wherein, S_tr1',S_tr2',…,S_trk',…,S_trN' represent respectively through asking the test obtained after binary image data to train Collection S_tr' in first training sample, second training sample ..., k-th training sample ..., n-th training sample, N is Number of training；

(3-4) gives training set S_tr' in sample from different type room set different labels, and constitute label matrix T；

(4) to radar range finding test set S_teSample information carry out feature extraction, concrete processing procedure is as follows：

(4-1) note test sample collection S_teIn any one training sample be S_J, 1≤J≤M, S_JIt it is one by radar scanning one week The one-dimensional characteristic vector that the radar data of acquisition is constituted, i.e. S_J=[S_J.1, S_J.2..., S_J.l], wherein S_J.1, S_J.2..., S_J.l The l radar data of sampled point in single pass is represented, it is converted into polar coordinate image by this group of radar data；

(4-2) extracts polar coordinate image in rectangular coordinate system, and using the center of circle of polar coordinate image as in rectangular image The heart, red filling is carried out to contoured interior point；

(4-3) carries out binary conversion treatment to the red blank map of the rectangle with white as background obtained in (4-2), becomes black White image, obtains view data, obtains the sample S of new training set_J', complete the feature extraction to training set sample；

And then obtain new test sample collection S_te'：

S_te'={ S_te1',S_te2',…,S_tek',…,S_teM'}；

Wherein, S_te1',S_te2',…,S_tek',…,S_teM' represent respectively through seeking the test set S obtained after binary image data_te' In first test sample, second test sample ..., k-th test sample ..., m-th test sample, M is test specimens This number；

(4-4) gives test set S in the way of in (3-4)_te' in sample from different type room set different labels, and Composition label matrix T'；

(5) by training set S_tr' and corresponding label matrix T, test set S_te' and corresponding label matrix T' as based on local sense It is input into by wild very fast learning machine, and the relevant parameters such as convolutional layer, pond layer is set；

(5-1) generates input weight W at random_i=[w_i1,,,w_in]^TWith the biasing b of Hidden unit_i=[b_i1,…b_in]^T, to initial Weight is orthogonalized, and obtains new input weightIf training set sample input size is (d × d), receptive field size is (r × r), k-th value c of convolution node (i, j) of characteristic pattern_i,j,kCalculated by following formula：

$c_{i,j,k}=\underset{m-1}{\overset{r}{\Σ}}\underset{n-1}{\overset{r}{\Σ}}(x_{i+m-1,j+n-1}\·a_{m,n,k}),i,j=1,...(d-r+1)$

(5-2) carries out square root pond to characteristic pattern, and pond size e represents Chi Hua centers to the distance on side, h_p,q,kRepresent k-th Combined joint (p, q) in Chi Huatu, is calculated as follows：

$h_{p,q,k}=\sqrt{\underset{i=p-e}{\overset{p+e}{\Σ}}\underset{j=q-e}{\overset{q+e}{\Σ}}{c}_{i,j,k}^2},p,q=1,...,(d-r+1);$

(5-3) simply connect all combined joints value formed a row vector, and N number of training set input sample row to Amount is put together, obtains combination layer matrix H, and output weight beta is calculated by following formula：

As N ＞ K (d-r+1)²,

$\mathrm{\β}={(\frac{I}{C}+H^{T}H)}^{-1}{H}^{T}T$

As N≤K (d-r+1)²,

$\mathrm{\β}=H^T{(\frac{I}{C}+{\mathrm{HH}}^T)}^{-1}T$

(5-4) input weightIt is constant, to test set S_te' sample carry out and identical convolution in (5-1) and (5-2) and pond Process, obtains combination layer H', and imputation method is Y to the Tag Estimation of test set sample, be can be calculated：

Y=H'* β

Prediction label Y is contrasted with the given label T' of test set, the accuracy of scene Recognition is drawn.