CN105608444A - Wild animal image identification method used for automatic driving - Google Patents

Abstract

The invention relates to a wild animal image identification method used for automatic driving. The method comprises the following steps: performing manifold learning on an acquired image by use of an LLE algorithm to obtain an image feature; training a nerve network through the image feature; and substituting an image acquired in real time into the nerve network for identification, obtaining a wild animal type label, and notifying a driver of a wild animal type corresponding to the label. According to the invention, a manifold learning idea is integrated into the method, and through obtaining low-dimensional expression of data, the original feature of the data can be better represented. Hidden layer nodes of the nerve network are solved by use of a K-means algorithm, due to a set c-type initial center, the iteration frequency of a K-means algorithm is substantially reduced, the time consumption of the algorithm is reduced, and the operation speed is fast.

Description

A kind of wild animal image-recognizing method for automatic Pilot

Technical field

The present invention relates to wild animal image-recognizing method, is a kind of wild moving for automatic Pilot specificallyObject image recognition methods.

Background technology

Along with the high speed development of human civilization, the coverage of mankind's activity is also accelerating expansion, along with highwayConstantly build, increasing highway has been built in the scope of activities of wild animal, many wild animalsBe faced with survival pressure be exactly and the mankind fight for living space, in our life, often can see this typeConflict. On Mou Tiao highway, the unexpected jaywalk of wild animal, the driver of automobile is expert at and crossesIn journey, do not expect, cause unnecessary injury. All a kind of loss for the mankind and the earth.

Along with automotive performance promotes, automobile has made the mankind's scope of activities more next as a kind of instrument of riding instead of walkMore next, even threatened the daily life of animals, and the high speed of automobile is for animal and Yan ShixiangWhen danger, animal cannot effectively be hidden the automobile of telling highway, how to driver's road that steps up vigilanceThe appearance of upper wild animal is a problem always.

Existing technology mostly, for pedestrian's detection, does not also have at present for the detection of wild animal, and movesThing and people's difference is the diversity of animal, and manifold learning arithmetic is seldom made the extraction effect of featureUse image processing techniques.

Summary of the invention

For above-mentioned technical deficiency, object of the present invention provides a kind of wild animal image for automatic PilotRecognition methods.

The technical solution adopted for the present invention to solve the technical problems is: a kind of wild moving for automatic PilotObject image recognition methods, comprises the following steps:

Adopt LLE algorithm to carry out manifold learning to gathering image, obtain characteristics of image;

By characteristics of image, neutral net is trained;

The image substitution neutral net of Real-time Collection is identified, obtained wild animal kind label, and willThe wild animal kind driver that this label is corresponding.

Described employing LLE algorithm carries out manifold learning to collection image and comprises the following steps:

The figure using image as sample architecture k-neighbour, and calculate similarity between any two images as closelySeemingly geodesic curve distance:

min(d_G(i,j),d_G(i,k)+d_G(k,j))

Wherein, d_GFor k-neighbour schemes the Euclidean distance between upper any two images, image index i, j,K is 1,2 ..., N, wherein N is image number;

Structural matrix M=(I-W)^T(I-W), wherein I is N × N unit matrix, and W is the nearly k-of N × NAdjacent figure matrix, k-neighbour schemes the approximate geodesic curve distance matrix between upper any two images;

Metzler matrix is carried out to feature decomposition, and X gets front m the characteristic vector of M as the result of feature extraction,Be characteristics of image X1 ... Xm.

Describedly by characteristics of image, neutral net is trained and is comprised the following steps:

Using characteristics of image as input, type of vehicle label is output, and hidden node is that K-means algorithm is poly-The center that class obtains;

Neutral net is trained to the weight that obtains the each node output of hidden layer.

Described K-means algorithm comprises the following steps:

1) select the initial center of c class: c is number of samples some/mono-, first sample is for severalJu Ji center, c sample is front c-1 the data point of distance point farthest in all data points; WhereinData set is X, and data point represents certain characteristics of image;

2) to any one sample, ask its distance that arrives c center, this sample is grouped into apart from the shortestThe class at heart place;

3) point in each class is averaged as such cluster centre; Return to step 2), until work asBefore till the cluster centre of all classes and the last iteration difference that obtains the cluster centre of all classes is less than threshold value.

The described image substitution neutral net by Real-time Collection is identified and is comprised the following steps:

Adopt LLE algorithm to carry out manifold learning to gathering image the image of Real-time Collection, obtain characteristics of image;

By characteristics of image and weight substitution neutral net, obtain wild animal kind label:

Y_NFor type of vehicle label, W is weight, D_NFor the distance matrix of current sample and each cluster centre, DFor the distance matrix between all cluster centres, p is hidden node number.

The present invention has following beneficial effect and advantage:

1. the present invention, by the thought of manifold learning, obtains the expression of image in low-dimensional, thus more effective generationThe feature of table original sample, for follow-up study and training provide the obvious characteristic of sample.

2. the present invention has incorporated the thought of manifold learning, express by the low-dimensional of obtaining data, thereby betterThe feature that representative data is original.

3. the present invention adopts K-means algorithm to ask for the hidden node of neutral net, and due to the c class of settingInitial center, the iterations of K-means algorithm is significantly reduced, reduce algorithm time loss, computing speedDegree is fast.

Brief description of the drawings

Fig. 1 is method flow diagram of the present invention;

Fig. 2 is neutral net schematic diagram of the present invention.

Detailed description of the invention

Below in conjunction with embodiment, the present invention is described in further detail.

Along with the rise of manifold learning, high dimensional data can better be expressed higher-dimension the potential stream shape of low-dimensionalFeature, the dimension that can find out image pattern of the pixel of image, can pass through manifold learning like thisAlgorithm realize the extraction of animal painting feature.

The low-dimensional that the present invention obtains various wild animal images by manifold learning LLE algorithm is expressed, as studyThe feature of machine is learnt learning machine, and object of the present invention makes running car arrive a certain region, intelligent vehicleLoading system can identify the wild animal that lives in this region appearing on front side road automatically, fromAnd dodge to driver with enough attentions, concrete step is as follows.

As shown in Figure 1, concrete step is as follows:

For all known type wild animals (the wild animal kind of zones of different is different, forThe wild animal kind that different region is detected is also different) feature extraction of image set, by all kind open countryLively thing is that main wild animal picture carries out feature extraction, is cut into the picture of fixed size, does with imageFor sample architecture k-neighbour figure, calculate the similarity matrix of any two pictures by calculating Euclidean distance:

Picture is lined up to vector according to row, and the distance between two vectors is constructed as the similarity of imageNeighbour figure, in the present invention, neighbour's number is total sample number 5%, for each sample its apart from minimumFront k is the neighbour of this sample. The front k width of every width picture and this picture are linked to be to a limit structure neighbour figure,Wherein in neighbour figure, there is the limit W on limit_ijValue is 1, and what all the other were boundless is 0, calculates institute's being similar between a little the shortestGeodesic curve distance:

min(d_G(i,j),d_G(i,k)+d_G(k,j))

d_GFor the Euclidean distance between any two images, i, j, k=1,2 ..., N, wherein N is that image is openedNumber; According to the distance between the shortest calculating of the distance between a figure mid point point, finally obtain the D of sample data collectionDistance matrix, wherein i, j=1,2 ..., N, wherein N is number of samples.

2. structure M=(I-W)^T(I-W), wherein I is N × N unit matrix, and W is N × N neighbour figure squareBattle array, carries out feature decomposition to Metzler matrix, and X gets a front M characteristic vector of M as the result of feature extraction, obtainsBe that the M dimension of data set is expressed.

3. as shown in Figure 2, RBF neural network model is learnt, and Y is carried out to k-means cluster, first to instituteThe feature that has image to obtainCarry out cluster (as the X1 in Fig. 2 ... Xm), M=m is input layerNumber, uses K-means algorithm to realize:

3-1 selects the initial center of c class: c is number of samples 1/10th, and first is data setCenter, first point of distance point farthest in second data point that is all formations, the 3rd is data setMiddle apart from the first two point distance point farthest ... by that analogy. Wherein data set is X, and data point represents certain figurePicture feature.

3-2, in the k time iteration, to any one sample, asks its distance that arrives c center, by this sampleBe grouped into the class apart from place, Duan center;

3-3 utilizes and calculates all points in such and average as the cluster centre that upgrades such;

3-4 is for all c cluster centre, if after utilizing (3-2) iterative method (3-3) to upgrade,Difference is less than 0.01, and iteration finishes, otherwise continues iteration.

Difference is the poor of the cluster centre of current all classes and the last iteration cluster centre that obtains all classes.

Training sample: by the barycenter { C obtaining₁，C₂,…,C_qAs the hidden node of neutral net, in Fig. 2's

Obtain sample as center by centre indexing:

Calculate the distance B D that arrives a little barycenter;

Design factor σ

$\mathrm{\σ}=\frac{D}{p*p-p}$

Wherein p is hidden layer number, and D is for arriving a little the distance of barycenter.

The input of calculating neuron

By what obtainIndependent a line that element of vector interpolation is 1 is as RBF's (radial base neural net)Coefficient

Design factor matrix is weights W

The label that wherein Y is sample is the Y in Fig. 2.

Also to there is the output of response for the vehicle of other types of not paying close attention to of output of result, unifiedly be output as itHis type.

4. according to the parameter obtaining, new samples is predicted

Calculate the Distance matrix D of target sample and barycenter C_N；

Calculate input coefficient

Added elements is that 1 row form RBF coefficient, and RBF coefficient is multiplied by weight and obtains the tag along sort of target

The Y finally obtaining_NBe the label of new samples.

The region in the process of travelling, the road in image being covered, as target area, is fixed at cameraIn time, is selected, and Real time identification is carried out in target area, sends police if be found to be known sample typeReport reminds driver to note, driver makes further deceleration again and can significantly reduce wild movingThe injury of thing.

Claims (5)

1. for a wild animal image-recognizing method for automatic Pilot, it is characterized in that comprising the following steps:

Adopt LLE algorithm to carry out manifold learning to gathering image, obtain characteristics of image;

By characteristics of image, neutral net is trained;

The image substitution neutral net of Real-time Collection is identified, obtained wild animal kind label, and willThe wild animal kind driver that this label is corresponding.

2. a kind of wild animal image-recognizing method for automatic Pilot according to claim 1, its featureBe that described employing LLE algorithm carries out manifold learning to collection image and comprises the following steps:

The figure using image as sample architecture k-neighbour, and calculate similarity between any two images as closelySeemingly geodesic curve distance:

min(d_G(i,j),d_G(i,k)+d_G(k,j))

Wherein, d_GFor k-neighbour schemes the Euclidean distance between upper any two images, image index i, j,K is 1,2 ..., N, wherein N is image number;

Structural matrix M=(I-W)^T(I-W), wherein I is N × N unit matrix, and W is the nearly k-of N × NAdjacent figure matrix, k-neighbour schemes the approximate geodesic curve distance matrix between upper any two images;

Metzler matrix is carried out to feature decomposition, and X gets front m the characteristic vector of M as the result of feature extraction,Be characteristics of image X1 ... Xm.

3. a kind of wild animal image-recognizing method for automatic Pilot according to claim 1, its featureDescribed in being, by characteristics of image, neutral net is trained and is comprised the following steps:

Using characteristics of image as input, type of vehicle label is output, and hidden node is that K-means algorithm is poly-The center that class obtains;

Neutral net is trained to the weight that obtains the each node output of hidden layer.

4. a kind of wild animal image-recognizing method for automatic Pilot according to claim 3, its featureBe that described K-means algorithm comprises the following steps:

1) select the initial center of c class: c is number of samples some/mono-, first sample is for severalJu Ji center, c sample is front c-1 the data point of distance point farthest in all data points; WhereinData set is X, and data point represents certain characteristics of image;

2) to any one sample, ask its distance that arrives c center, this sample is grouped into apart from the shortestThe class at heart place;

3) point in each class is averaged as such cluster centre; Return to step 2), until work asBefore till the cluster centre of all classes and the last iteration difference that obtains the cluster centre of all classes is less than threshold value.

5. a kind of wild animal image-recognizing method for automatic Pilot according to claim 1, its featureBeing that the described image substitution neutral net by Real-time Collection is identified comprises the following steps:

Adopt LLE algorithm to carry out manifold learning to gathering image the image of Real-time Collection, obtain characteristics of image;

By characteristics of image and weight substitution neutral net, obtain wild animal kind label:

$\mathrm{\σ}=\frac{D}{p*p-p}$

Y_NFor wild animal kind label, W is weight, D_NFor the distance matrix of current sample and each cluster centre,D is the distance matrix between all cluster centres, and p is hidden node number.