A kind of classification of landform method based on coorinated training
Technical field
The present invention relates to robot fields, more particularly to a kind of classification of landform method based on coorinated training.
Background technique
Different from doors structure environment when working in robot outer scene out of office, landform-ground in the scene of field is more
Sample, complexity can produce bigger effect the travelling performance of robot.Accurate classification of the robot to terrain environment locating for it
It is to determine that can it realize the key factor of autonomous.Tactile and vision are the mankind in the judgement for carrying out locating terrain environment
When common perceptive mode, therefore we attempt to divide robot also using similar perceptual model to locating terrain environment
Class.Wherein tactile can only realize the perception for now locating landform to robot, and larger range of landform perception may be implemented in vision, right
The landform that will be travelling through can equally be perceived.In the present invention, we by using the haptic data that has mark on a small quantity and
Vision data makes full use of the data not marked largely, the mode of two classifier coorinated training is taken to carry out the instruction of classifier
Practice, to reduce dependence of the classifier training process to labeled data, and realizes the ground that will be travelling through ground to robot simultaneously
The judgement of shape classification.
Summary of the invention
The technology of the present invention overcome the deficiencies in the prior art solves and marks missing in view-based access control model and the classification of landform of vibration
The training problem of situation.
To solve the above problems, the invention discloses a kind of classification of landform method based on coorinated training, specifically include with
Lower step:
Step S1: enabling robot, traveling for a period of time, is collected simultaneously machine respectively in every kind of landform in its working environment
The ground of the camera record of the haptic signal time series and face forward ground of the touch sensor output of Qi Ren foot installation
Face image sequence is split the haptic signal time series acquired in every kind of landform respectively, and segmentation length is α sampling
Point obtains the set of the corresponding haptic signal tract of every kind of landform;Feature extraction is carried out to each haptic signal tract, is obtained
To the corresponding haptic signal sample set of every kind of landform, the set of the haptic signal sample set of all landform is denoted asFeature extraction is carried out to the image in the corresponding ground image sequence of every kind of landform, obtains every kind of ground
The corresponding characteristics of image sample set of shape, the set of the characteristics of image sample set of all landform are denoted asIts
Middle aιAnd bιIndicate the haptic signal sample that touch sensor of the robot on same place obtains and the image that video camera obtains
Feature samples, Indicate that haptic signal sample set A and characteristics of image sample set B includes the quantity of sample;
With landform number 1,2, J is labeled the sample in these sample sets, and wherein J indicates the sum of terrain type, obtains
To there is mark sample set { A, B, Ψ },For terrain type set corresponding with A and B, wherein ψι∈
{ 1,2, J } is indicated and aιAnd bιCorresponding terrain type;
Step S2: enabling robot random walk in its working environment, walked in the haptic signal time sequence collected
Column and ground image sequence, are split haptic signal time series, and segmentation length is α sampled point, obtain in walking
The set of the haptic signal tract of collection carries out feature extraction to haptic signal tract, obtains haptic signal sample set, remembers
ForFeature extraction is carried out to the image in the ground image sequence collected in walking, obtains image spy
Sample set is levied, is denoted asWherein c κ and dκIndicate that touch sensor of the robot on same place obtains
The characteristics of image sample that the haptic signal sample and video camera taken obtains, Indicate haptic signal sample
Collect the quantity that C and characteristics of image sample set D includes sample;Thus it obtains without mark sample set { C, D };
Step S3: enabling repetitive exercise serial number e=0, remembers the training set of the classifier based on haptic signal and is based on surface map
The training set of the classifier of picture is respectively L(1)、L(2), and enable L(1)={ A, Ψ }, L(2)={ B, Ψ };
Step S4: it is based on training set L(1)And L(2)Classifier based on haptic signal is respectively trained and based on ground image
Classifier, sorter model use support vector machines, remember that trained classifier isWith
Step S5: n is randomly selected from haptic signal sample set CcA sample is input to classifierIn, it respectively obtains
ncThe set of the landform prediction result of a sampleAnd the confidence level of landform prediction resultThe wherein highest n of confidence levelc′A landform prediction result is used to mark corresponding characteristics of image
Sample forms markd characteristics of image sample set V(2);N is randomly selected from characteristics of image sample set DdA sample is input to
ClassifierIn, respectively obtain ndThe set of the landform prediction result of a sampleAnd landform
The confidence level of prediction resultThe wherein highest n of confidence leveld′A landform prediction result is used to mark
Corresponding haptic signal sample forms markd haptic signal sample set V(1);By V(1)And V(2)It is added separately toWithTraining set in, i.e. L(1)←L(1)∪V(1), L(2)←L(2)∪V(2);It will set V(1)Corresponding nc′A haptic signal sample
This is deleted from C, by set V(2)Corresponding nd′A characteristics of image sample is deleted from D, i.e.,;WhereinIndicate set V(1)Corresponding nc′A haptic signal sample composition
Set,Indicate set V(2)Corresponding nd′The set of a characteristics of image sample composition;
Step S6: enabling e ← e+1, if e is less than preset repetitive exercise frequency threshold value T, repeatedly step 4 and 5;It is no
Then, classifier training terminates, and obtains final classifier C(1)And C(2);When robot in actual work, by the surface map of acquisition
After the operation Jing Guo feature extraction, it is input to classifier C(2)In, obtained output is robot to the ground that will be passed through
The prediction result of shape type.
Compared with existing technology, the invention has the following advantages that
Category forecasting is carried out to the landform that will be travelling through 1. can realize, so that robot can change traveling strategy in advance,
Reduce the adaptation time to topographic change;
2. classifier is respectively trained using two kinds of perceptive modes, and the mode of coorinated training is carried out, makes full use of and do not mark
The accuracy rate of data raising classifier;
3. since training process is smaller to the dependence for having labeled data, the time required to labeled data can be greatlyd save.
Detailed description of the invention
Fig. 1 is flow chart of the present invention.
Specific embodiment
In order to make the objectives, technical solutions, and advantages of the present invention clearer, below in conjunction with attached drawing and specific implementation
The present invention is described in detail for example.
A kind of classification of landform method based on coorinated training, as shown in Figure 1, specifically includes the following steps:
Step S1: enabling robot, traveling for a period of time, is collected simultaneously machine respectively in every kind of landform in its working environment
The ground of the camera record of the haptic signal time series and face forward ground of the touch sensor output of Qi Ren foot installation
Face image sequence is split the haptic signal time series acquired in every kind of landform respectively, and segmentation length is α sampling
Point obtains the set of the corresponding haptic signal tract of every kind of landform;Feature extraction is carried out to each haptic signal tract, is obtained
To the corresponding haptic signal sample set of every kind of landform, the set of the haptic signal sample set of all landform is denoted asFeature extraction is carried out to the image in the corresponding ground image sequence of every kind of landform, obtains every kind of ground
The corresponding characteristics of image sample set of shape, the set of the characteristics of image sample set of all landform are denoted asIts
Middle aιAnd bιIndicate the haptic signal sample that touch sensor of the robot on same place obtains and the image that video camera obtains
Feature samples, Indicate that haptic signal sample set A and characteristics of image sample set B includes the quantity of sample;
With landform number 1,2, J is labeled the sample in these sample sets, and wherein J indicates the sum of terrain type, obtains
To there is mark sample set { A, B, Ψ },For terrain type set corresponding with A and B, wherein ψι∈
{ 1,2, J } indicate and aιAnd bιCorresponding terrain type;
Step S2: enabling robot random walk in its working environment, walked in the haptic signal time sequence collected
Column and ground image sequence, are split haptic signal time series, and segmentation length is α sampled point, obtain in walking
The set of the haptic signal tract of collection carries out feature extraction to haptic signal tract, obtains haptic signal sample set, remembers
ForFeature extraction is carried out to the image in the ground image sequence collected in walking, obtains image spy
Sample set is levied, is denoted asWherein cκAnd dκIndicate that touch sensor of the robot on same place obtains
The characteristics of image sample that the haptic signal sample and video camera taken obtains, Indicate haptic signal sample
Collect the quantity that C and characteristics of image sample set D includes sample;Thus it obtains without mark sample set { C, D };
Step S3: enabling repetitive exercise serial number e=0, remembers the training set of the classifier based on haptic signal and is based on surface map
The training set of the classifier of picture is respectively L(1)、L(2), and enable L(1)={ A, Ψ }, L(2)={ B, Ψ };
Step S4: it is based on training set L(1)And L(2)Classifier based on haptic signal is respectively trained and based on ground image
Classifier, sorter model use support vector machines, remember that trained classifier isWith
Step S5: n is randomly selected from haptic signal sample set CcA sample is input to classifierIn, it respectively obtains
ncThe set of the landform prediction result of a sampleAnd the confidence level of landform prediction resultThe wherein highest n of confidence levelc′A landform prediction result is used to mark corresponding characteristics of image
Sample forms markd characteristics of image sample set V(2);N is randomly selected from characteristics of image sample set DdA sample is input to
ClassifierIn, respectively obtain ndThe set of the landform prediction result of a sampleAnd landform
The confidence level of prediction resultThe wherein highest n of confidence leveld′A landform prediction result is used to mark
Corresponding haptic signal sample forms markd haptic signal sample set V(1);By V(1)And V(2)It is added separately toWithTraining set in, i.e. L(1)←L(1)∪V(1), L(2)←L(2)∪V(2);It will set V(1)Corresponding nc′A haptic signal sample
This is deleted from C, by set V(2)Corresponding nd′A characteristics of image sample is deleted from D, i.e.,;WhereinIndicate set V(1)Corresponding nc′A haptic signal sample composition
Set,Indicate set V(2)Corresponding nd′The set of a characteristics of image sample composition;
Step S6: enabling e ← e+1, if e is less than preset repetitive exercise frequency threshold value T, repeatedly step 4 and 5;It is no
Then, classifier training terminates, and obtains final classifier C(1)And C(2);When robot in actual work, by the surface map of acquisition
After the operation Jing Guo feature extraction, it is input to classifier C(2)In, obtained output is robot to the ground that will be passed through
The prediction result of shape type.
In implementing process of the present invention, the feature extraction side of a variety of haptic signal tracts and ground image can be used
Method.For example, for haptic signal tract Fast Fourier Transform (FFT) can be carried out to it, point of its amplitude in frequency is obtained
These amplitudes are arranged in vector according to the sequence of frequency from small to large by cloth, that is, the feature for realizing haptic signal tract mentions
It takes.For ground image, its color histogram can be extracted, the frequency of different colours is arranged in vector, that is, realizes ground
The feature extraction of image.
Above embodiments are provided just for the sake of the description purpose of the present invention, and are not intended to limit the scope of the invention.This
The range of invention is defined by the following claims.It does not depart from spirit and principles of the present invention and the various equivalent replacements made and repairs
Change, should all cover within the scope of the present invention.