CN104268503B - pedestrian detection method and device - Google Patents

Abstract

The invention relates to a pedestrian detection method and device, and belongs to the technical field of computer vision. The method includes: obtaining pedestrian samples and non-pedestrian samples respectively; obtaining a pedestrian sample dictionary and a non-pedestrian sample dictionary through the K-SVD algorithm; obtaining the residual error of the sample to be tested in the pedestrian sample dictionary and the non-pedestrian sample dictionary; When the residual error of the sample in the pedestrian sample dictionary is smaller than the residual error in the non-pedestrian sample dictionary, it is determined that the sample to be tested contains pedestrians. The present invention trains pedestrian samples and non-pedestrian samples into two types of dictionaries respectively through the K-SVD algorithm, and judges whether pedestrians are contained in the samples to be tested according to the size relationship of the residual errors of the samples to be tested in the two types of dictionaries; When the technology is applied to environments with occlusion and illumination changes, there is a problem of high false alarm rate, and the effect of pedestrian detection can be effectively applied to environments with occlusion and illumination changes.

Description

Pedestrian detection method and device

technical field

The invention relates to the technical field of computer vision, in particular to a pedestrian detection method and device.

Background technique

In recent years, pedestrian detection has attracted more and more attention, and pedestrian detection in images and videos can be used in areas such as intelligent surveillance, intelligent transportation, and motion analysis. At the same time, people have more and more requirements for pedestrian detection.

There is a pedestrian detection method in the related art. Firstly, the AdaBoost (Adaptive Boosting, self-adaptive enhancement) detection classifier is trained by the calculated target features. The block detects whether a pedestrian is present in this image.

In the process of realizing the present invention, the inventor found that the above-mentioned method has at least the following defects: the above-mentioned method cannot describe the image information from multiple angles through the fixed feature extraction process, and the training process of the classifier is also relatively complicated. False positive rates are higher in environments with occlusions and lighting changes.

Contents of the invention

In order to solve the problem that image information cannot be described from multiple angles in the related art, the training process of the classifier is also relatively complicated, and when it is applied to an environment with occlusion and illumination changes, the problem of a high false alarm rate, the embodiment of the present invention provides Provided a pedestrian detection method and device, the technical solution is as follows:

According to a first aspect of an embodiment of the present invention, a pedestrian detection method is provided, the method comprising:

Obtain pedestrian samples and non-pedestrian samples respectively;

The pedestrian sample is trained as a pedestrian sample dictionary by K-singular value decomposition K-SVD algorithm;

Using the K-SVD algorithm to train the non-pedestrian samples into a dictionary of non-pedestrian samples;

Acquiring the residual error of the sample to be tested in the pedestrian sample dictionary, the sample to be tested is a sample that needs to be detected by pedestrians;

Obtain the residual error of the sample to be tested in the dictionary of non-pedestrian samples;

When the residual error of the sample to be tested in the dictionary of pedestrian samples is smaller than the residual error of the sample to be tested in the dictionary of non-pedestrian samples, it is determined that the sample to be tested contains a pedestrian.

Optionally, said pedestrian sample training is a dictionary of pedestrian samples through the K-singular value decomposition K-SVD algorithm, including:

Converting the pedestrian samples into pedestrian edge feature samples;

The pedestrian edge feature sample is trained as a pedestrian sample dictionary by the K-SVD algorithm;

Said training said non-pedestrian sample as a non-pedestrian sample dictionary through said K-SVD algorithm includes:

Converting the non-pedestrian samples into non-pedestrian edge feature samples;

The non-pedestrian edge feature samples are trained as a non-pedestrian sample dictionary through the K-SVD algorithm.

Optionally, using the K-SVD algorithm to train the pedestrian edge feature samples into a pedestrian sample dictionary includes:

By solving the training equation, the pedestrian edge feature sample is trained as a pedestrian sample dictionary;

The non-pedestrian edge feature sample is trained as a non-pedestrian sample dictionary by the K-SVD algorithm, including:

By solving the training equation, the non-pedestrian edge feature sample is trained as a non-pedestrian sample dictionary;

Wherein said training equation is:

The constraints of the training equation are: ||x _j || ₀ ≤ L; where D _i represents a dictionary, y _ij represents the pedestrian edge feature sample or non-pedestrian edge feature sample, i=0 or 1, when i=0, D _i represents the non-pedestrian sample dictionary, y _ij represents non-pedestrian edge feature samples, when i=1, D _i represents pedestrian sample dictionary, y _ij represents pedestrian edge feature samples, x _j represents sparse coding, 1≤j≤K and j is an integer, L is the sparsity control coefficient .

Optionally, the obtaining the residual error of the sample to be tested in the pedestrian sample dictionary, the sample to be tested is a sample that requires pedestrian detection, including:

Obtaining the pedestrian sparse coding of the sample to be tested in the pedestrian sample dictionary by solving the sparse coding equation, the sample to be tested is a sample that requires pedestrian detection;

calculating the residual error of the sample to be tested in the pedestrian sample dictionary according to the pedestrian sparse coding;

The acquisition of the residual error of the sample to be tested in the non-pedestrian sample dictionary includes:

Obtaining the non-pedestrian sparse coding of the sample to be tested in the non-pedestrian sample dictionary by solving the sparse coding equation;

calculating the residual error of the sample to be tested in the non-pedestrian sample dictionary according to the non-pedestrian sparse coding;

Wherein, the sparse coding equation is:

x _i ＝argmin||x _i || ₀

The constraints of the sparse coding equation are: D _i x _i =q, i=0 or 1; where _xi represents sparse coding, D _i represents a dictionary, when i=0, D _i represents a dictionary of non-pedestrian samples, and _xi represents For non-pedestrian sparse coding, when i=1, D _i represents the pedestrian sample dictionary, x _i represents the pedestrian sparse coding, and q represents the edge feature sample of the sample to be tested.

Optionally, the calculating the residual error of the sample to be tested in the pedestrian sample dictionary according to the pedestrian sparse coding includes:

Obtaining the residual error of the sample to be tested in the pedestrian sample dictionary by solving the residual error equation;

The calculation of the residual error of the sample to be tested in the non-pedestrian sample dictionary according to the non-pedestrian sparse coding includes:

Obtaining the residual error of the sample to be tested in the non-pedestrian sample dictionary by solving the residual error equation;

The residual error equation is:

r _i (y)＝||qD _i x _i || ₂

Among them, r _i (y) represents the residual error, i=0 or 1, when i=0, r _i (y) represents the residual error of the sample to be tested in the non-pedestrian sample dictionary, and when i=1, r _i (y) represents the residual error of the test sample in the pedestrian sample dictionary.

According to a second aspect of an embodiment of the present invention, a pedestrian detection device is provided, the device comprising:

The sample acquisition module is used to obtain pedestrian samples and non-pedestrian samples respectively;

The pedestrian dictionary module is used to train the pedestrian sample as a pedestrian sample dictionary by the K-singular value decomposition K-SVD algorithm;

A non-pedestrian dictionary module, configured to train the non-pedestrian samples into a non-pedestrian sample dictionary by the K-SVD algorithm;

Pedestrian error module, used to obtain the residual error of the sample to be tested in the pedestrian sample dictionary, the sample to be tested is a sample that needs to be detected by pedestrians;

A non-pedestrian error module, configured to obtain the residual error of the sample to be tested in the non-pedestrian sample dictionary;

a pedestrian determination module, configured to determine that the sample to be tested contains pedestrian.

Optionally, the pedestrian dictionary module includes:

A pedestrian feature unit, configured to convert the pedestrian sample into a pedestrian edge feature sample;

A pedestrian training unit, configured to train the pedestrian edge feature samples as a pedestrian sample dictionary through the K-SVD algorithm;

The non-pedestrian dictionary module includes:

A non-pedestrian feature unit, configured to convert the non-pedestrian sample into a non-pedestrian edge feature sample;

The non-pedestrian training unit is configured to use the K-SVD algorithm to train the non-pedestrian edge feature samples into a non-pedestrian sample dictionary.

Optionally, the pedestrian training unit is specifically configured to train the pedestrian edge feature samples into a pedestrian sample dictionary by solving a training equation;

The non-pedestrian training unit is specifically used to train the non-pedestrian edge feature samples into a non-pedestrian sample dictionary by solving a training equation;

Wherein, the training equation is:

The constraints of the training equation are: ||x _j || ₀ ≤ L; where D _i represents a dictionary, y _ij represents the pedestrian edge feature sample or non-pedestrian edge feature sample, i=0 or 1, when i=0, D _i represents the non-pedestrian sample dictionary, y _ij represents non-pedestrian edge feature samples, when i=1, D _i represents pedestrian sample dictionary, y _ij represents pedestrian edge feature samples, x _j represents sparse coding, 1≤j≤K and j is an integer, L is the sparsity control coefficient .

Optionally, the pedestrian error module includes:

a pedestrian encoding unit, configured to obtain the pedestrian sparse encoding of the sample to be tested in the pedestrian sample dictionary by solving a sparse encoding equation, and the sample to be tested is a sample that requires pedestrian detection;

a pedestrian error unit, configured to calculate the residual error of the sample to be tested in the pedestrian sample dictionary according to the pedestrian sparse coding;

The non-pedestrian error module includes:

A non-pedestrian encoding unit, configured to obtain the non-pedestrian sparse encoding of the sample to be tested in the non-pedestrian sample dictionary by solving a sparse encoding equation;

A non-pedestrian error unit, configured to calculate the residual error of the sample to be tested in the non-pedestrian sample dictionary according to the non-pedestrian sparse coding;

Wherein, the sparse coding equation is:

x _i ＝argmin||x _i || ₀

The constraints of the sparse coding equation are: D _i x _i =q, i=0 or 1; where _xi represents sparse coding, D _i represents a dictionary, when i=0, D _i represents a dictionary of non-pedestrian samples, and _xi represents For non-pedestrian sparse coding, when i=1, D _i represents the pedestrian sample dictionary, x _i represents the pedestrian sparse coding, and q represents the edge feature sample of the sample to be tested.

Optionally, the pedestrian error unit is specifically configured to obtain the residual error of the sample to be tested in the pedestrian sample dictionary by solving a residual error equation;

The non-pedestrian error unit is specifically used to obtain the residual error of the sample to be tested in the non-pedestrian sample dictionary by solving a residual error equation;

The residual error equation is:

r _i (y)＝||qD _i x _i || ₂

Among them, r _i (y) represents the residual error, i=0 or 1, when i=0, r _i (y) represents the residual error of the sample to be tested in the non-pedestrian sample dictionary, and when i=1, r _i (y) represents the residual error of the test sample in the pedestrian sample dictionary.

The technical solutions provided by the embodiments of the present invention may include the following beneficial effects:

Through the K-SVD algorithm, the pedestrian samples and non-pedestrian samples are trained into two types of dictionaries, and according to the size relationship of the residual errors of the samples to be tested in the two types of dictionaries, it is judged whether there are pedestrians in the samples to be tested; solves the problem in related technologies It is impossible to describe the image information from multiple angles, and the training process of the classifier is also relatively complicated. When it is applied to an environment with occlusion and illumination changes, the problem of high false alarm rate has reached the matching degree of the image as a whole. It is detected and can be effectively applied to the effect of pedestrian detection in environments with occlusion and illumination changes.

It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the invention.

Description of drawings

The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and together with the description serve to explain the principles of the invention.

Fig. 1 is a flowchart of a pedestrian detection method shown according to an exemplary embodiment;

Fig. 2 is a flow chart of a pedestrian detection method according to another exemplary embodiment;

Fig. 3 is a block diagram of a pedestrian detection device according to an exemplary embodiment;

Fig. 4 is a block diagram of a pedestrian detection device according to another exemplary embodiment.

By way of the above drawings, specific embodiments of the invention have been shown and will be described in more detail hereinafter. These drawings and written descriptions are not intended to limit the scope of the inventive concept in any way, but to illustrate the inventive concept for those skilled in the art by referring to specific embodiments.

Detailed ways

Reference will now be made in detail to the exemplary embodiments, examples of which are illustrated in the accompanying drawings. When the following description refers to the accompanying drawings, the same numerals in different drawings refer to the same or similar elements unless otherwise indicated. The implementations described in the following exemplary examples do not represent all implementations consistent with the present invention. Rather, they are merely examples of apparatuses and methods consistent with aspects of the invention as recited in the appended claims.

Fig. 1 is a flow chart of a pedestrian detection method according to an exemplary embodiment. This embodiment is described by taking the pedestrian detection method applied to a terminal as an example. The terminal in the embodiment of the present invention may be any one of a mobile phone, a tablet computer, a laptop computer, a driving recorder, a camera, a desktop computer, and a video camera. The pedestrian detection method may include the following steps:

In step 101, pedestrian samples and non-pedestrian samples are obtained respectively.

In step 102, the pedestrian samples are trained as a dictionary of pedestrian samples through the K-singular value decomposition K-SVD algorithm.

In step 103, the non-pedestrian samples are trained into a dictionary of non-pedestrian samples through the K-SVD algorithm.

In step 104, the residual error of the sample to be tested in the dictionary of pedestrian samples is obtained, and the sample to be tested is a sample that needs to be detected for pedestrians.

In step 105, the residual error of the sample to be tested in the non-pedestrian sample dictionary is obtained.

In step 106, when the residual error of the sample to be tested in the pedestrian sample dictionary is smaller than the residual error of the sample to be tested in the non-pedestrian sample dictionary, it is determined that the sample to be tested contains a pedestrian.

In summary, the pedestrian detection method provided in this embodiment uses the K-SVD algorithm to train pedestrian samples and non-pedestrian samples into two types of dictionaries respectively, and according to the relationship between the residual errors of the samples to be tested in the two types of dictionaries, Judging whether there are pedestrians in the sample to be tested; solving the problem that the image information cannot be described from multiple angles in related technologies, and the training process of the classifier is also relatively complicated. When applied to environments with occlusion and illumination changes, the false positive rate is relatively high The high problem has achieved the detection of the matching degree of the image as a whole, and can be effectively applied to the effect of pedestrian detection in environments with occlusion and illumination changes.

Fig. 2 is a flow chart showing a pedestrian detection method according to another exemplary embodiment. This embodiment is described by taking the pedestrian detection method applied to a terminal as an example. The pedestrian detection method may include the following steps:

In step 201, the terminal acquires pedestrian samples and non-pedestrian samples respectively.

When the terminal detects pedestrians, it first obtains pedestrian samples and non-pedestrian samples respectively, where pedestrian samples and non-pedestrian samples can contain the same predetermined number of images, that is, pedestrian samples contain a predetermined number of pedestrian images, and non-pedestrian samples contain Book number of non-pedestrian images. Preferably, the subscription quantity can be selected from 600 to 2000.

Wherein the non-pedestrian sample can be a predetermined number of non-pedestrian images that do not contain pedestrians, such as a predetermined number of images that do not contain pedestrians collected by a driving recorder; pedestrian samples can be a predetermined number of pedestrian images that include a pedestrian, such as collected by a driving recorder A predetermined number of images of a pedestrian are included.

In step 202, the terminal converts pedestrian samples into pedestrian edge feature samples.

After acquiring the pedestrian samples, the terminal converts the acquired pedestrian samples into pedestrian edge feature samples.

When the pedestrian samples are a predetermined number of pedestrian images, the following takes converting one of the images into an edge feature sample as an example to illustrate the conversion process in this step:

1) Scale the pedestrian image to a predetermined size and convert it into a grayscale image.

2) Use the Sobel operator to calculate the gradient of each pixel in the grayscale image.

where the Sobel operator can be:

The gradient of each pixel in the x direction is g _x , the gradient in the y direction is g _y , g _x =t*s _x , g _y =t*s _y , t can be the 8 pixels around the pixel grayscale value.

3) The terminal calculates the gradient magnitude of the pixel point according to the gradient.

Gradient magnitude M(x,y)=|g _x |+|g _y |.

4) After calculating the gradient magnitudes of all pixels in the grayscale image, use the gradient magnitudes to form the edge feature map of the grayscale image, and then convert the edge feature map into a dimension vector to obtain a pedestrian edge feature sample.

When the pedestrian samples include a predetermined number of pedestrian images, each of the pedestrian images is processed as 1) to 4) to obtain a predetermined number of pedestrian edge feature samples.

In step 203, the terminal uses the K-SVD algorithm to train pedestrian edge feature samples into a dictionary of pedestrian samples.

After obtaining the predetermined number of pedestrian edge feature samples, the terminal trains the predetermined number of pedestrian edge feature samples into a pedestrian sample dictionary through the K-SVD algorithm.

Among them, the terminal can train the pedestrian edge feature samples into a pedestrian sample dictionary by solving the training equation, and the training equation is:

The constraints of the training equation are: ||x _j || ₀ ≤ L; where D _i represents a dictionary, y _ij represents a pedestrian edge feature sample or a non-pedestrian edge feature sample, i=0 or 1, when i=0, D _i represents a non-pedestrian sample dictionary, y _ij Represents non-pedestrian edge feature samples, when i=1, D _i represents pedestrian sample dictionary, y _ij represents pedestrian edge feature samples, x _j represents sparse coding, 1≤j≤K and j is an integer, j is the number of samples, L is Sparsity control coefficient, which can be set by the user, such as 3 or 4. The purpose of this training equation is to find the D _i that is the closest to D _i x _j and y _ij when there are few non-zero elements in x _j , and the obtained D _i can be expressed as an n-dimensional column vector, n Represents the size of the dictionary, ||y _ij -D _i x _j || ₂ represents the L2 norm of y _ij -D _i x _j .

That is, the terminal can solve the training equation by using the pedestrian edge feature samples y _1j to obtain the pedestrian sample dictionary D ₁ .

The above training equation is a sparse coding equation, which can be solved by the orthogonal matching pursuit algorithm, because using the orthogonal matching pursuit algorithm to solve the sparse coding equation is a common technical means for those skilled in the art, so it will not be described here. In addition, other commonly used There are also technical means such as base tracking algorithm and so on.

In step 204, the terminal converts the non-pedestrian samples into non-pedestrian edge feature samples.

This step is basically the same as the method in step 202, and will not be described here.

In step 205, the terminal uses the K-SVD algorithm to train non-pedestrian edge feature samples into a dictionary of non-pedestrian samples.

This step is basically the same as the method in step 203, that is, the terminal obtains the non-pedestrian sample dictionary D ₀ by solving the training equation through the non-pedestrian edge feature sample y _0j .

It should be noted that the order of the steps of the pedestrian detection method provided by the embodiment of the present invention can be adjusted appropriately, and the steps can also be increased or decreased according to the situation. For example, step 204 and step 205 can also be before step 202 and step 203 That is, the pedestrian detection method provided in this embodiment only needs to perform step 203 after step 202, and step 205 after step 204. Any person familiar with the technical field can easily think of The changing methods should all be included in the protection scope of the present invention, so no more details are given here.

In step 206, the terminal obtains the pedestrian sparse coding of the sample to be tested in the pedestrian sample dictionary by solving the sparse coding equation, and the sample to be tested is a sample that requires pedestrian detection.

This step can include the following three sub-steps:

1) The terminal obtains the sample to be tested.

The sample to be tested may be an image acquired by the terminal, such as an image in front of the vehicle captured in real time by a driving recorder.

2) The terminal converts the samples to be tested into edge feature samples.

For occlusion, the samples to be tested can be transformed into edge feature samples q by the method in step 202 .

3) The terminal obtains the sparse coding of the edge feature sample in the pedestrian sample dictionary by solving the sparse coding equation.

The terminal solves the sparse coding equation through the edge feature sample q, and obtains the sparse coding of the sample to be tested in the pedestrian sample dictionary.

Among them, the sparse coding equation is:

x _i ＝argmin||x _i || ₀

The constraints of the sparse coding equation are: D _i x _i =q, i=0 or 1; where _xi represents sparse coding, D _i represents a dictionary, when i=0, D _i represents a dictionary of non-pedestrian samples, and _xi represents a non-pedestrian Sparse coding, when i=1, D _i represents the pedestrian sample dictionary, _xi represents pedestrian sparse coding, q represents the edge feature sample of the sample to be tested, and || _xi || ₀ represents the L0 norm of _xi .

That is, the pedestrian sparse coding x ₁ of the sample to be tested in the pedestrian sample dictionary can be obtained through the sparse coding equation.

In step 207, the terminal calculates the residual error of the sample to be tested in the pedestrian sample dictionary according to the pedestrian sparse coding.

The terminal obtains the residual error of the sample to be tested in the pedestrian sample dictionary by solving the residual error equation.

The residual error equation is:

r _i (y)＝||qD _i x _i || ₂

Among them, ri (y) represents the residual error, _i =0 or 1, when _i =0, ri (y) represents the residual error of the sample to be tested in the non-pedestrian sample dictionary, and when _i =1, ri (y) represents The residual error of the tested sample in the pedestrian sample dictionary.

That is, the terminal can obtain the residual error r ₁ (y) of the sample to be tested in the pedestrian sample dictionary through the error equation.

In step 208, the terminal obtains the non-pedestrian sparse coding of the sample to be tested in the non-pedestrian sample dictionary by solving the sparse coding equation.

This step is basically the same as the method in step 206, that is, the non-pedestrian sparse coding of the sample to be tested in the non-pedestrian sample dictionary can be obtained through the method in step 206.

In step 209, the terminal calculates the residual error of the sample to be tested in the non-pedestrian sample dictionary according to the non-pedestrian sparse coding.

This step is basically the same as the method in step 207, that is, the residual error r ₀ (y) of the sample to be tested in the non-pedestrian sample dictionary can be obtained by the method in step 207.

It should be noted that the sequence of the steps of the pedestrian detection method provided by the embodiment of the present invention can be appropriately adjusted, and the steps can also be increased or decreased according to the situation. For example, step 208 and step 209 can also be before step 206 and step 207 That is, the pedestrian detection method provided by this embodiment only needs to carry out step 209 after step 208, step 208 is carried out after step 205, step 207 is carried out after step 206, and step 206 is carried out after step 203, anyone familiar with the technology Those skilled in the art within the technical scope disclosed in the present invention can easily think of changing methods, all of which should be covered within the protection scope of the present invention, so details are not repeated here.

In step 210, when the residual error of the sample to be tested in the dictionary of pedestrian samples is smaller than the residual error of the sample to be tested in the dictionary of non-pedestrian samples, the terminal determines that the sample to be tested contains pedestrians.

The terminal judges the size relationship between r ₀ (y) and r ₁ (y). When r ₀ (y) > r ₁ (y), the terminal determines that the sample to be tested contains pedestrians. When r ₀ (y) ≤ r ₁ (y), the terminal determines that the sample to be tested does not include pedestrians.

It should be added that the pedestrian detection method provided in this embodiment uses edge feature samples to train the pedestrian sample dictionary and non-pedestrian sample dictionary, and the key to identifying objects depends on the edges of objects, so this method can effectively reduce background interference .

It should be added that the Adaboost detection classifier in the prior art needs to set parameters such as sample weight and error rate threshold, the size of the sliding window and the step size during detection also need to be set, and the pedestrian provided by this embodiment The detection method only needs to simply set a sparsity control coefficient L, which achieves the effect of making the structure of the pedestrian detection method simple and easy to use in actual scenes.

It should be added that the pedestrian detection method provided in this embodiment is to determine whether there are pedestrians in the sample to be tested by training two types of dictionaries for pedestrians and non-pedestrians, and by comparing the residual errors of the samples to be tested in the two types of dictionaries. It has the effect of increasing the accuracy rate of pedestrian detection method.

It needs to be added that each column vector in the two types of dictionaries trained by the pedestrian detection method provided in this embodiment can be expressed as a primitive that composes an image, and the object and occlusion can be separated on different primitives, achieving The effect of improving the performance of the pedestrian detection method of this embodiment in the case of occlusion.

In summary, the pedestrian detection method provided in this embodiment uses the K-SVD algorithm to train pedestrian samples and non-pedestrian samples into two types of dictionaries respectively, and according to the relationship between the residual errors of the samples to be tested in the two types of dictionaries, Judging whether there are pedestrians in the sample to be tested; solving the problem that the image information cannot be described from multiple angles in related technologies, and the training process of the classifier is also relatively complicated. When applied to environments with occlusion and illumination changes, the false positive rate is relatively high The high problem has achieved the detection of the matching degree of the image as a whole, and can be effectively applied to the effect of pedestrian detection in environments with occlusion and illumination changes.

The following are device embodiments of the present invention, which can be used to implement the method embodiments of the present invention. For the details not disclosed in the device embodiment of the present invention, please refer to the method embodiment of the present invention.

Fig. 3 is a block diagram of a pedestrian detection device according to an exemplary embodiment. The pedestrian detection device can be implemented as part or all of a terminal by software, hardware or a combination of the two. The pedestrian detection system may include: a sample acquisition module 310, a pedestrian dictionary module 320, a non-pedestrian dictionary module 330, a pedestrian error module 340, a non-pedestrian error module 350 and a pedestrian determination module 360;

A sample acquisition module 310, configured to acquire pedestrian samples and non-pedestrian samples respectively;

The pedestrian dictionary module 320 is used to train the pedestrian sample as a pedestrian sample dictionary through the K-singular value decomposition K-SVD algorithm;

The non-pedestrian dictionary module 330 is used to train non-pedestrian samples into a non-pedestrian sample dictionary by the K-SVD algorithm;

The pedestrian error module 340 is used to obtain the residual error of the sample to be tested in the pedestrian sample dictionary, and the sample to be tested is a sample that needs to be detected by pedestrians;

The non-pedestrian error module 350 is used to obtain the residual error of the sample to be tested in the non-pedestrian sample dictionary;

The pedestrian determining module 360 is configured to determine that the sample to be tested contains a pedestrian when the residual error of the sample to be tested in the dictionary of pedestrian samples is smaller than the residual error of the sample to be tested in the dictionary of non-pedestrian samples.

In summary, the pedestrian detection device provided in this embodiment uses the K-SVD algorithm to train pedestrian samples and non-pedestrian samples into two types of dictionaries respectively, and according to the relationship between the residual errors of the samples to be tested in the two types of dictionaries, Judging whether there are pedestrians in the sample to be tested; solving the problem that the image information cannot be described from multiple angles in related technologies, and the training process of the classifier is also relatively complicated. When applied to environments with occlusion and illumination changes, the false positive rate is relatively high The high problem has achieved the detection of the matching degree of the image as a whole, and can be effectively applied to the effect of pedestrian detection in environments with occlusion and illumination changes.

Fig. 4 is a block diagram of a pedestrian detection device according to another exemplary embodiment. The pedestrian detection device may be implemented as part or all of a terminal by software, hardware or a combination of the two. The pedestrian detection system may include: a sample acquisition module 310, a pedestrian dictionary module 320, a non-pedestrian dictionary module 330, a pedestrian error module 340, a non-pedestrian error module 350 and a pedestrian determination module 360;

A sample acquisition module 310, configured to acquire pedestrian samples and non-pedestrian samples respectively;

The pedestrian dictionary module 320 is used to train the pedestrian sample as a pedestrian sample dictionary through the K-singular value decomposition K-SVD algorithm;

The non-pedestrian dictionary module 330 is used to train non-pedestrian samples into a non-pedestrian sample dictionary by the K-SVD algorithm;

The pedestrian error module 340 is used to obtain the residual error of the sample to be tested in the pedestrian sample dictionary, and the sample to be tested is a sample that needs to be detected by pedestrians;

The non-pedestrian error module 350 is used to obtain the residual error of the sample to be tested in the non-pedestrian sample dictionary;

The pedestrian determining module 360 is configured to determine that the sample to be tested contains a pedestrian when the residual error of the sample to be tested in the dictionary of pedestrian samples is smaller than the residual error of the sample to be tested in the dictionary of non-pedestrian samples.

Optionally, the pedestrian dictionary module 320 includes:

Pedestrian feature unit 321, for converting pedestrian samples into pedestrian edge feature samples;

Pedestrian training unit 322, used to train pedestrian edge feature samples as pedestrian sample dictionary by K-SVD algorithm;

Non-pedestrian dictionary module 330, comprising:

Non-pedestrian feature unit 331, for converting non-pedestrian samples into non-pedestrian edge feature samples;

The non-pedestrian training unit 332 is configured to train non-pedestrian edge feature samples into a dictionary of non-pedestrian samples through the K-SVD algorithm.

Optionally, the pedestrian training unit 322 is specifically configured to train pedestrian edge feature samples into a pedestrian sample dictionary by solving a training equation;

The non-pedestrian training unit 332 is specifically used to train the non-pedestrian edge feature samples into a non-pedestrian sample dictionary by solving the training equation;

Among them, the training equation is:

The constraints of the training equation are: ||x _j || ₀ ≤ L; where D _i represents a dictionary, y _ij represents a pedestrian edge feature sample or a non-pedestrian edge feature sample, i=0 or 1, when i=0, D _i represents a non-pedestrian sample dictionary, y _ij Represents non-pedestrian edge feature samples, when i=1, D _i represents pedestrian sample dictionary, y _ij represents pedestrian edge feature samples, x _j represents sparse coding, 1≤j≤K and j is an integer, L is the sparsity control coefficient.

Optionally, the pedestrian error module 340 includes:

The pedestrian encoding unit 341 is used to obtain the pedestrian sparse encoding of the sample to be tested in the pedestrian sample dictionary by solving the sparse encoding equation, and the sample to be tested is a sample that requires pedestrian detection;

Pedestrian error unit 342, for calculating the residual error of the sample to be tested in the pedestrian sample dictionary according to the pedestrian sparse coding;

A non-pedestrian error module 350 comprising:

The non-pedestrian coding unit 351 is used to obtain the non-pedestrian sparse coding of the sample to be tested in the non-pedestrian sample dictionary by solving the sparse coding equation;

The non-pedestrian error unit 352 is used to calculate the residual error of the sample to be tested in the non-pedestrian sample dictionary according to the non-pedestrian sparse coding;

Among them, the sparse coding equation is:

x _i ＝argmin||x _i || ₀

The constraints of the sparse coding equation are: D _i x _i =q, i=0 or 1; where _xi represents sparse coding, D _i represents a dictionary, when i=0, D _i represents a dictionary of non-pedestrian samples, and _xi represents a non-pedestrian Sparse coding, when i=1, D _i represents the pedestrian sample dictionary, x _i represents the pedestrian sparse coding, and q represents the edge feature sample of the sample to be tested.

Optionally, the pedestrian error unit 342 is specifically configured to obtain the residual error of the sample to be tested in the pedestrian sample dictionary by solving the residual error equation;

The non-pedestrian error unit 352 is specifically used to obtain the residual error of the sample to be tested in the non-pedestrian sample dictionary by solving the residual error equation;

The residual error equation is:

r _i (y)＝||qD _i x _i || ₂

Among them, ri (y) represents the residual error, _i =0 or 1, when _i =0, ri (y) represents the residual error of the sample to be tested in the non-pedestrian sample dictionary, and when _i =1, ri (y) represents The residual error of the tested sample in the pedestrian sample dictionary.

In summary, the pedestrian detection device provided in this embodiment uses the K-SVD algorithm to train pedestrian samples and non-pedestrian samples into two types of dictionaries respectively, and according to the relationship between the residual errors of the samples to be tested in the two types of dictionaries, Judging whether there are pedestrians in the sample to be tested; solving the problem that the image information cannot be described from multiple angles in related technologies, and the training process of the classifier is also relatively complicated. When applied to environments with occlusion and illumination changes, the false positive rate is relatively high The high problem has achieved the detection of the matching degree of the image as a whole, and can be effectively applied to the effect of pedestrian detection in environments with occlusion and illumination changes.

Regarding the apparatus in the foregoing embodiments, the specific manner in which each module executes operations has been described in detail in the embodiments related to the method, and will not be described in detail here.

Other embodiments of the invention will be readily apparent to those skilled in the art from consideration of the specification and practice of the invention invented herein. This application is intended to cover any modification, use or adaptation of the present invention, these modifications, uses or adaptations follow the general principles of the present invention and include common knowledge or conventional technical means in the technical field not invented by the present invention . The specification and examples are to be considered exemplary only, with a true scope and spirit of the invention being indicated by the following claims.

It should be understood that the present invention is not limited to the precise constructions which have been described above and shown in the accompanying drawings, and various modifications and changes may be made without departing from the scope thereof. The scope of the invention is limited only by the appended claims.

Claims (4)

1. A pedestrian detection method, characterized in that the method comprises: Obtain pedestrian samples and non-pedestrian samples respectively; The pedestrian sample is trained as a pedestrian sample dictionary by K-singular value decomposition K-SVD algorithm; Using the K-SVD algorithm to train the non-pedestrian samples into a dictionary of non-pedestrian samples; Acquiring the residual error of the sample to be tested in the pedestrian sample dictionary, the sample to be tested is a sample that needs to be detected by pedestrians; Obtain the residual error of the sample to be tested in the dictionary of non-pedestrian samples; When the residual error of the sample to be tested in the pedestrian sample dictionary is smaller than the residual error of the sample to be tested in the non-pedestrian sample dictionary, it is determined that the sample to be tested contains a pedestrian; The pedestrian sample is trained as a pedestrian sample dictionary by the K-singular value decomposition K-SVD algorithm, including: Converting the pedestrian samples into pedestrian edge feature samples; The pedestrian edge feature sample is trained as a pedestrian sample dictionary by the K-SVD algorithm; Said training said non-pedestrian sample as a non-pedestrian sample dictionary through said K-SVD algorithm includes: Converting the non-pedestrian samples into non-pedestrian edge feature samples; The non-pedestrian edge feature sample is trained as a non-pedestrian sample dictionary by the K-SVD algorithm; The described pedestrian edge feature sample is trained as a pedestrian sample dictionary through the K-SVD algorithm, including: By solving the training equation, the pedestrian edge feature sample is trained as a pedestrian sample dictionary; The non-pedestrian edge feature sample is trained as a non-pedestrian sample dictionary by the K-SVD algorithm, including: By solving the training equation, the non-pedestrian edge feature sample is trained as a non-pedestrian sample dictionary; Wherein, the training equation is: <mrow><msub><mi>D</mi><mi>i</mi></msub><mo>=</mo><mi>arg</mi><munder><mrow><mi>m</mi><mi>i</mi><mi>n</mi></mrow><msub><mi>x</mi><mi>j</mi></msub></munder><mrow><mo>(</mo><munderover><mo>&amp;Sigma;</mo><mrow><mi>j</mi><mo>=</mo><mn>1</mn></mrow><mi>K</mi></munderover><mo>|</mo><mo>|</mo><msub><mi>y</mi><mrow><mi>i</mi><mi>j</mi></mrow></msub><mo>-</mo><msub><mi>D</mi><mi>i</mi></msub><msub><mi>x</mi><mi>j</mi></msub><mo>|</mo><msubsup><mo>|</mo><mn>2</mn><mn>2</mn></msubsup><mo>)</mo></mrow></mrow> The constraints of the training equation are: Where D _i represents a dictionary, y _ij represents the pedestrian edge feature sample or non-pedestrian edge feature sample, i=0 or 1, when i=0, D _i represents the non-pedestrian sample dictionary, y _ij represents the non-pedestrian edge feature sample, i =1, D _i represents the pedestrian sample dictionary, y _ij represents the pedestrian edge feature samples, x _j represents sparse coding, 1≤j≤K and j is an integer, L is the sparsity control coefficient; The acquisition of the residual error of the sample to be tested in the pedestrian sample dictionary, the sample to be tested is a sample that requires pedestrian detection, including: Obtaining the pedestrian sparse coding of the sample to be tested in the pedestrian sample dictionary by solving the sparse coding equation, the sample to be tested is a sample that requires pedestrian detection; calculating the residual error of the sample to be tested in the pedestrian sample dictionary according to the pedestrian sparse coding; The acquisition of the residual error of the sample to be tested in the non-pedestrian sample dictionary includes: Obtaining the non-pedestrian sparse coding of the sample to be tested in the non-pedestrian sample dictionary by solving the sparse coding equation; calculating the residual error of the sample to be tested in the non-pedestrian sample dictionary according to the non-pedestrian sparse coding; Wherein, the sparse coding equation is: x _i ＝argmin||x _i || ₀ The constraints of the sparse coding equation are: D _i x _i =q, i=0 or 1; where _xi represents sparse coding, D _i represents a dictionary, when i=0, D _i represents a dictionary of non-pedestrian samples, and _xi represents For non-pedestrian sparse coding, when i=1, D _i represents the pedestrian sample dictionary, x _i represents the pedestrian sparse coding, and q represents the edge feature sample of the sample to be tested. 2. The method of claim 1, wherein, The calculating the residual error of the sample to be tested in the pedestrian sample dictionary according to the pedestrian sparse coding includes: Obtaining the residual error of the sample to be tested in the pedestrian sample dictionary by solving the residual error equation; The calculation of the residual error of the sample to be tested in the non-pedestrian sample dictionary according to the non-pedestrian sparse coding includes: Obtaining the residual error of the sample to be tested in the non-pedestrian sample dictionary by solving the residual error equation; The residual error equation is: r _i (y)＝||qD _i x _i || ₂ Among them, r _i (y) represents the residual error, i=0 or 1, when i=0, r _i (y) represents the residual error of the sample to be tested in the non-pedestrian sample dictionary, and when i=1, r _i (y) represents the residual error of the test sample in the pedestrian sample dictionary. 3. A pedestrian detection device, characterized in that the device comprises: The sample acquisition module is used to obtain pedestrian samples and non-pedestrian samples respectively; The pedestrian dictionary module is used to train the pedestrian sample as a pedestrian sample dictionary through the K-singular value decomposition K-SVD algorithm, and the pedestrian dictionary module includes: A pedestrian feature unit, configured to convert the pedestrian sample into a pedestrian edge feature sample; A pedestrian training unit, configured to train the pedestrian edge feature samples as a pedestrian sample dictionary through the K-SVD algorithm; The non-pedestrian dictionary module is used to train the non-pedestrian sample into a non-pedestrian sample dictionary by the K-SVD algorithm, and the non-pedestrian dictionary module includes: A non-pedestrian feature unit, configured to convert the non-pedestrian sample into a non-pedestrian edge feature sample; A non-pedestrian training unit, configured to train the non-pedestrian edge feature samples as a non-pedestrian sample dictionary through the K-SVD algorithm; Pedestrian error module, used to obtain the residual error of the sample to be tested in the pedestrian sample dictionary, the sample to be tested is a sample that needs to be detected by pedestrians; A non-pedestrian error module, configured to obtain the residual error of the sample to be tested in the non-pedestrian sample dictionary; a pedestrian determination module, configured to determine that the sample to be tested contains pedestrian; The pedestrian training unit is specifically configured to train the pedestrian edge feature samples into a pedestrian sample dictionary by solving a training equation; The non-pedestrian training unit is specifically used to train the non-pedestrian edge feature samples into a non-pedestrian sample dictionary by solving a training equation; Wherein, the training equation is: <mrow><msub><mi>D</mi><mi>i</mi></msub><mo>=</mo><mi>arg</mi><munder><mrow><mi>m</mi><mi>i</mi><mi>n</mi></mrow><msub><mi>x</mi><mi>j</mi></msub></munder><mrow><mo>(</mo><munderover><mo>&amp;Sigma;</mo><mrow><mi>j</mi><mo>=</mo><mn>1</mn></mrow><mi>K</mi></munderover><mo>|</mo><mo>|</mo><msub><mi>y</mi><mrow><mi>i</mi><mi>j</mi></mrow></msub><mo>-</mo><msub><mi>D</mi><mi>i</mi></msub><msub><mi>x</mi><mi>j</mi></msub><mo>|</mo><msubsup><mo>|</mo><mn>2</mn><mn>2</mn></msubsup><mo>)</mo></mrow></mrow> The constraints of the training equation are: Where D _i represents a dictionary, y _ij represents the pedestrian edge feature sample or non-pedestrian edge feature sample, i=0 or 1, when i=0, D _i represents the non-pedestrian sample dictionary, y _ij represents the non-pedestrian edge feature sample, i =1, D _i represents the pedestrian sample dictionary, y _ij represents the pedestrian edge feature samples, x _j represents sparse coding, 1≤j≤K and j is an integer, L is the sparsity control coefficient; The pedestrian error module includes: a pedestrian encoding unit, configured to obtain the pedestrian sparse encoding of the sample to be tested in the pedestrian sample dictionary by solving a sparse encoding equation, and the sample to be tested is a sample that requires pedestrian detection; a pedestrian error unit, configured to calculate the residual error of the sample to be tested in the pedestrian sample dictionary according to the pedestrian sparse coding; The non-pedestrian error module includes: A non-pedestrian encoding unit, configured to obtain the non-pedestrian sparse encoding of the sample to be tested in the non-pedestrian sample dictionary by solving a sparse encoding equation; A non-pedestrian error unit, configured to calculate the residual error of the sample to be tested in the non-pedestrian sample dictionary according to the non-pedestrian sparse coding; Wherein, the sparse coding equation is: x _i ＝argmin||x _i || ₀ The constraints of the sparse coding equation are: D _i x _i =q, i=0 or 1; where _xi represents sparse coding, D _i represents a dictionary, when i=0, D _i represents a dictionary of non-pedestrian samples, and _xi represents For non-pedestrian sparse coding, when i=1, D _i represents the pedestrian sample dictionary, x _i represents the pedestrian sparse coding, and q represents the edge feature sample of the sample to be tested. 4. The device of claim 3, wherein: The pedestrian error unit is specifically used to obtain the residual error of the sample to be tested in the pedestrian sample dictionary by solving a residual error equation; The non-pedestrian error unit is specifically used to obtain the residual error of the sample to be tested in the non-pedestrian sample dictionary by solving a residual error equation; The residual error equation is: r _i (y)＝||qD _i x _i || ₂ Among them, r _i (y) represents the residual error, i=0 or 1, when i=0, r _i (y) represents the residual error of the sample to be tested in the non-pedestrian sample dictionary, and when i=1, r _i (y) represents the residual error of the test sample in the pedestrian sample dictionary.