CN104778453B - A kind of night pedestrian detection method based on infrared pedestrian's brightness statistics feature - Google Patents

Abstract

The invention discloses a kind of night pedestrian detection method based on infrared pedestrian's brightness statistics feature.The present invention counts gray average information to each part of pedestrian in Sample Storehouse and negative sample first, and mapping range border is determined using the information, constructs a brightness histogram feature for distinguishing ballot interval division；Then gradient orientation histogram feature is calculated, and the two features are subjected to joint and form final feature descriptor；Secondly model training is carried out using the method for Adaboost combination decision trees, and pedestrian's judgement and positioning are carried out by sliding window scanning method, finally when grader is obtained compared with low confidence to the classification judgement of some detection block, judged again using brightness section template, so as to realize the pedestrian detection at night.The present invention have effectively achieved the pedestrian detection under night environment, have the characteristics of high verification and measurement ratio, strong adaptability.

Description

A Pedestrian Detection Method at Night Based on Infrared Pedestrian Brightness Statistical Features

technical field

The invention relates to a pedestrian detection method of a vehicle-mounted video image, in particular to a pedestrian detection method at night based on the statistical characteristics of infrared pedestrian brightness.

Background technique

Pedestrian detection technology is an important application of computer vision, which has high practical value in daily life and production. Pedestrian detection based on vision is to judge the specific location of pedestrians from the input picture or video frame sequence according to certain image processing technology. The intelligent vehicle assisted driving system can improve the safety of vehicle driving, thereby reducing the occurrence of traffic accidents, and pedestrian detection technology is one of the core technologies in the intelligent assisted driving system.

Vision-based pedestrian detection technology at night mainly uses visible light images, infrared images and other technologies. At night, due to unsatisfactory lighting and other conditions, the imaging effect of the visible light camera is poor, which affects the effect of pedestrian detection. Infrared cameras capture infrared information to perceive objects through passive infrared technology, and objects with different temperatures show different brightness in the image. In the infrared image of the road scene, pedestrians generally radiate more heat than the background, and the pedestrians in the infrared image are generally brighter than the background, and are not affected by dark light at night, unclear vision in foggy days, etc., and have good night vision ability , have strong adaptability to different lighting environments. Therefore, the pedestrian detection technology of infrared images is an effective solution to realize pedestrian detection at night.

At present, most infrared pedestrian detection technologies use machine learning-based methods. For example, the invention patent "Pedestrian Detection Method and Device Based on Feature Combination" (CN103632170A) uses HOG features and LBP features to form a joint feature, and uses support vector machines (SVM) as learning Algorithm for classifier training to achieve pedestrian detection. The LBP feature involved in the above method has a strong ability to describe the image texture, but the texture feature of the infrared image is not obvious, so the effect of the LBP feature applied to infrared pedestrian detection is general. Invention patent "A Pedestrian Detection Method Based on Infrared Image" (CN103902976A) based on HOG feature fusion and brightness histogram features to describe pedestrians, using SVM for classifier training to realize nighttime pedestrian detection. In addition to using the outline information of pedestrians, this method also extracts the brightness information of infrared pedestrians, so it has a better detection effect of infrared pedestrians.

The above methods do not consider the statistical characteristics of the brightness distribution of infrared pedestrians. Based on the HOG feature, the present invention combines the statistical characteristics of the brightness distribution of infrared pedestrians to construct an infrared pedestrian feature with stronger description ability, uses Adaboost as a learning algorithm, and designs a pedestrian detection classifier and a pedestrian framework for re-determining the brightness interval template . The method of the invention has better pedestrian feature description ability and detection effect in infrared pedestrian detection.

Contents of the invention

The purpose of the present invention is to provide a pedestrian detection method at night based on statistical characteristics of infrared pedestrian brightness. The present invention utilizes the profile information and brightness information of infrared pedestrians, firstly calculates the average gray value information of the pedestrian parts (including head, upper body, lower body) and negative sample images in the sample library, and determines the pedestrian according to the distribution characteristics of the gray average value information Describe each voting mapping interval of the feature, construct a brightness histogram feature (DBHOI) that distinguishes the voting interval division; then calculate the gradient orientation histogram (HOG) feature, and combine these two features to form the final pedestrian feature descriptor; Secondly, use Adaboost combined with the decision tree method for model training, and use the sliding window scanning method to determine and locate pedestrians; finally, when the classifier obtains a low confidence level for the classification and judgment of a certain pedestrian frame, it uses the brightness interval template for re-judgment , so as to realize infrared pedestrian detection at night. This method has the characteristics of high detection rate and strong adaptability.

The technical solution steps adopted by the present invention to solve its technical problems are as follows:

Step 1. Construct positive and negative sample data sets of infrared images;

Step 2, infrared image DBHOI feature construction based on brightness statistical features;

Step 3, infrared image HOG feature construction based on contour information;

Step 4, utilize Adaboost to carry out classifier training;

Step 5, detection window judgment and positioning;

Step 6: Pedestrian frame weight determination.

The details of constructing the positive and negative sample data sets of infrared images described in step 1 are as follows:

The method of constructing the positive sample data set is as follows: use the smallest rectangular window to extract pedestrian samples in the infrared image; assume that the height of the pedestrian is h and the width is w, so that w/h=0.41; a total of N ₁ positive samples of pedestrians are extracted;

The negative sample data set construction method is as follows: Randomly extract N ₀ × 10 negative samples from N ₀ infrared images that do not contain pedestrians, that is, randomly select 10 negative samples from each infrared image;

Scales all positive and negative samples to a sample image with a width of 64 pixels and a height of 128 pixels.

The infrared image DBHOI feature construction based on the brightness statistical feature described in step 2 is specifically as follows:

2-1. Set the size of the sample image as w′×h′, then divide the sample image into multiple partial images of the same size; the size of each partial image is 8×8, that is, divide the sample image into equal (w′ /8)×(h′/8) partial images, and record these partial images as cells;

2-2. Determine the mapping rules according to the pedestrian components;

2-2-1. Firstly, according to N ₁ positive samples, each part of the pedestrian is intercepted, and each part of the pedestrian includes the head, upper body, and lower body;

2-2-2. Randomly intercept N ₁ background images in the negative sample;

2-2-3. Calculate the gray mean value of the head image, upper body image, lower body image and background image respectively;

2-2-4. Set the average gray values of the four images as G ₁ , G ₂ , G ₃ , and G ₄ , and the sizes increase sequentially; and determine three mapping boundaries according to the four average gray values, respectively t ₁ =(G ₁ +G ₂ )/2, t ₂ =(G ₂ +G ₃ )/2, t ₃ =(G ₃ +G ₄ )/2, divide the gray value range into four gray values according to the mapping boundary degree interval, respectively [0,t ₁ ),[t ₁ ,t ₂ ),[t ₂ ,t ₃ ),[t ₃ ,255];

2-2-5. According to the mapping rule and using the gray value of the pixel in the infrared image as the weight, construct a brightness histogram in each cell to obtain a four-dimensional feature vector;

2-3. Adjacent 2×2 cells are normalized within the block;

Taking one cell as the step size, put (w'/8)×(h'/8) cells in the sample image in order from top to bottom and from left to right to pass the adjacent four cells through L1-Sqrt method for intra-block normalization;

The L1-Sqrt method is as follows: v is the brightness histogram vector, ε is a very small value, the value is 0.001;

2-4. Concatenate the features of all blocks to obtain the DBHOI features;

The DBHOI is a brightness histogram with different interval sizes.

The HOG feature structure of the infrared image based on the contour information described in step 3 is specifically as follows:

Use the classic sobel operator to calculate the gradient components G _x (i, j) and G _y (i, j) of each pixel in the horizontal and vertical directions, and then calculate the corresponding gradient size G (i, j) and direction D (i, j) as follows:

The classifier training using Adaboost described in step 4 is as follows:

4-1. Scale all positive and negative samples to the same scale, then extract the DBHOI feature vector and HOG feature vector for each sample image, and mark the label of the positive sample as 1, and the label of the negative sample as -1;

4-2. Input the DBHOI-HOG feature vectors and sample labels corresponding to N ₁ positive samples and N ₀ ×10 negative samples to the Adaboost learning algorithm for training, and obtain a classifier composed of a series of weak classifiers.

The determination and positioning of the detection window described in step 5 are as follows:

5-1. Determine the scaling factor according to the scale of the infrared image to be detected and the size of the detection window, and determine the number of scaling layers of the infrared image to be detected; then use the classifier obtained in step 4 to scan layer by layer with a step size of 4 pixels ; Suppose the original size of the infrared image to be detected is W _i ×H _i , where W _i represents the width of the infrared image to be detected, H _i represents the height of the infrared image to be detected, and the size of the detection window is W _d ×H _d , where W _d represents The width of the detection window, H _d represents the height of the detection window, and s _s represents the scaling factor; then the initial scaling factor is s _s =1, and the termination scaling factor is s _s =min{W _i /W _d ,H _i /H _d } ; Sliding window scanning is performed on the infrared image to be detected at each scale, and the detection window is judged by using the trained pedestrian classifier; if the detection window is a pedestrian frame, record the position and confidence of the detection window, and record the Expressed as {posX, posY, width, height, score}, where posX and posY are the upper left corner of the pedestrian frame, width and height are the width and height of the pedestrian frame, and score is the confidence level;

The confidence degree is the error rate recorded by all the leaf nodes of the two-layer decision tree;

5-2. After multi-scale sliding window scanning detection of the infrared image to be detected, the same pedestrian is detected in different scales of the infrared image to be detected, and the non-maximum value suppression method is used to detect multiple pedestrian frames of different scales. The results are merged;

The criterion for maximum value suppression is the confidence score of each pedestrian frame;

5-3. Arrange all the pedestrian frames into an array A[n] according to the confidence level from low to high, and then take out the detection window information A[n] with the highest current confidence level from the array A[n];

5-4. Determine the relationship between the pedestrian frame A[n] and the subsequent pedestrian frame A[n-1]. If the coincidence degree α of the two pedestrian frames is greater than 0.5, it is considered to be the same pedestrian frame, otherwise the pedestrian frame A [n-1] as the pedestrian frame with the highest current confidence;

The degree of coincidence Where area(B _n ) is the area of the nth pedestrian frame, area(B _n ∩B _n-1 ) is the intersection of the area of the nth pedestrian frame and the area of the n-1th pedestrian frame;

5-5. Repeat steps 5-4 until all pedestrian frames are determined.

The pedestrian frame weight determination described in step 6 is as follows:

6-1. After passing step 5, each detection window that is judged as a pedestrian will have a confidence value score; the pedestrian frame is re-judged using the brightness interval template; when the score _i ≤τ, the pedestrian frame is re-determined Judgment, where the value of the threshold τ is determined by a statistical method; if score _i > τ, only the classifier is used for pedestrian detection, assuming that the confidence of N pedestrian frames is obtained, then τ satisfies the formula

The score _i is the i-th confidence value in the array A[n] defined in step 5-3;

6-2. According to step 2 and the characteristics of infrared pedestrian imaging, the four intervals [0,t ₁ ),[t ₁ ,t ₂ ),[t ₂ ,t ₃ ),[t ₃ ,255] correspond to the background, Upper body, lower body, head; if the width and height of the pedestrian is w×h, if the pedestrian frame correctly determines the position of the pedestrian, then the height h of the pedestrian frame that has been judged as a pedestrian is divided into 4 equal parts from top to bottom, then at Brightness information belonging to the range [t ₃ ,255] must exist in a 1/4 equal division, and the total number of pixels whose gray value is in this interval is Σp ₁ , and the second 1/4 equal division must exist [t ₁ , t ₂ ) range of brightness information, and record the total number of grayscale pixels in this range as Σp ₂ , and the last two 1/4 equal parts must exist in the range of [t ₂ , t ₃ ) luminance information, and record the total number of pixels with gray values in this interval as Σp ₃ ; set the following verification rules: Σp ₁ /(w×h)≥1/16, Σp ₂ /(w×h)≥1/ 8. Σp ₃ /(w×h)≥1/16; if Σp ₁ , Σp ₂ , and Σp ₃ meet these three conditions at the same time, then the pedestrian frame is judged as a pedestrian frame again, otherwise the pedestrian frame is judged as a non-pedestrian frame. Beneficial effects of the present invention:

Aiming at the characteristics of pedestrians in infrared images, the present invention constructs a more expressive brightness feature descriptor——DBHOI, and combines it with HOG features. The DBHOI descriptor counts the brightness distribution of pedestrian parts and background in training samples, and encodes the distribution information when constructing features, so that the feature descriptor can better describe the difference in brightness information between pedestrians and background. The construction method of the feature descriptor improves the description ability of the feature, so that the final classifier has a stronger classification ability.

When the classifier is used to determine the pedestrian frame and get a lower confidence, the pedestrian frame is re-determined according to the brightness interval template. This method greatly reduces the error rate caused by the misjudgment of the classifier and improves the detection accuracy.

Description of drawings

Fig. 1 is the overall flow chart of the present invention.

Figure 2 is a flow chart of DBHOI feature descriptor extraction.

Fig. 3 is a display of the detection effect of the pedestrian detection model of the present invention.

detailed description

The specific embodiments of the present invention will be further described in detail below in conjunction with the accompanying drawings.

As shown in Figure 1, Figure 2 and Figure 3, a night pedestrian detection method based on the statistical characteristics of infrared pedestrian brightness, which specifically includes the following steps:

Step 1. Construct positive and negative sample datasets of infrared images.

The present invention is mainly aimed at the detection of pedestrians in the case of vehicles, so the scenes for collecting images are mainly infrared images in road scenes.

The method of constructing the positive sample data set is as follows: Use the smallest rectangular window, that is to say, use a rectangular frame that can just surround pedestrians to extract pedestrian samples in the infrared image; assume that the height of the pedestrian is h and the width is w, so that w/h = 0.41; A total of N ₁ positive pedestrian samples were extracted;

The negative sample data set construction method is as follows: Randomly extract N ₀ × 10 negative samples from N ₀ infrared images that do not contain pedestrians, that is, randomly select 10 negative samples from each infrared image;

Scales all positive and negative samples to a sample image with a width of 64 pixels and a height of 128 pixels.

The definition of the minimum rectangular window is as described by the author P. Dollár in the document "Pedestrian Detection: AnEvaluation of the State of the Art".

Step 2: Construction of infrared image DBHOI features based on brightness statistical features.

The DBHOI is a brightness histogram of different interval sizes, and its full name is Different Bins Histogram of Intensity.

As shown in FIG. 2 , the present invention encodes the brightness information in the infrared image, so that the brightness information becomes a descriptor with strong resolution capability, and its specific structure is as follows.

2-1. Assuming that the size of the sample image is w′×h′, the sample image is divided into multiple partial images of the same size; the size of each partial image is 8×8, that is, the sample image is divided into equal (w′ /8)×(h′/8) partial images, and record these partial images as cells.

2-2. Determine the mapping rule according to the pedestrian component. Firstly, according to N ₁ positive samples, each part of the pedestrian is intercepted, and each part of the pedestrian includes the head, upper body, and lower body; then N ₁ background images are randomly intercepted in the negative samples, and then the head image, upper body image, lower body image and background are calculated respectively The grayscale mean of the image. Assume that the average gray levels of the four images are G ₁ , G ₂ , G ₃ , and G ₄ , and the sizes increase sequentially; and determine three mapping boundaries according to the four gray average values, which are respectively t ₁ =(G ₁ +G ₂ )/2, t ₂ ＝(G ₂ +G ₃ )/2, t ₃ ＝(G ₃ +G ₄ )/2, according to the mapping boundary, the gray value range is divided into four gray ranges, which are [ 0,t ₁ ),[t ₁ ,t ₂ ),[t ₂ ,t ₃ ),[t ₃ ,255]. According to the mapping rule and with the gray value as the weight, a brightness histogram is constructed in each cell to obtain a four-dimensional feature vector.

2-3. Adjacent 2×2 cells are normalized within the block.

cells, with a cell as the step size, (w′/8)×(h′/8) cells are placed in the sample image in order from top to bottom and from left to right. Intra-block normalization. The normalization method is L1-Sqrt: Where v is the brightness histogram vector, ε is a very small value, which is 0.001 in the present invention.

2-4. Concatenate the features of all blocks to obtain the DBHOI features. The DBHOI descriptor constructed in this way has strong descriptive ability.

Step 3. The infrared image HOG feature construction based on the contour information.

Use the classic sobel operator to calculate the gradient components G _x (i, j) and G _y (i, j) of each pixel in the horizontal and vertical directions, and then calculate the corresponding gradient size G (i, j) and direction D (i, j) as follows:

The construction of HOG features adopts the construction method of Jindian, such as dividing the sample image of 64×128 size into several cells, and the size of each cell is 8×8. For each cell area, the gradient direction is divided into 9 equal parts, and the gradient size is used as the weight to construct a gradient direction histogram. The adjacent 4 cells are combined into blocks, and the L1-sqrt method is used for normalization, and finally the gradient histograms between different blocks are concatenated to form the final HOG descriptor.

Step 4: Use Adaboost for classifier training.

Scale all positive and negative samples to the same scale, such as 64×128. Then for each sample image, extract the DBHOI feature vector and HOG feature vector, and mark the label of the positive sample as 1, and the label of the negative sample as -1.

The DBHOI-HOG feature vectors and sample labels corresponding to N ₁ positive samples and N ₀ × 10 negative samples are input into the Adaboost learning algorithm for training, and a classifier consisting of a series of weak classifiers is obtained. Among them, the weak classifier is a two-layer decision tree model.

Step 5: Pedestrian frame determination and positioning.

As shown in Figure 3, since the size of the pedestrian classifier is fixed, the size of pedestrians in the infrared image is different. For example, when a pedestrian is very close to the camera, its size will be larger than that of the classifier. Therefore, in order to detect pedestrians at different scales, it is necessary to zoom the image, and then perform a window scan for the image at each scale. The same pedestrian may be judged as a pedestrian by the classifier at different image scales, so the fusion of pedestrian frames is required.

5-1. Determine the scaling factor according to the scale of the infrared image to be detected and the size of the detection window, and determine the number of scaling layers of the infrared image to be detected; then use the classifier obtained in step 4 to scan layer by layer with a step size of 4 pixels . Suppose the original size of the infrared image to be detected is W _i ×H _i , where W _i represents the width of the infrared image to be detected, H _i represents the height of the infrared image to be detected, and the size of the detection window is W _d ×H _d , where W _d represents the detection The width of the window, H _d represents the height of the detection window, and s _s represents the scaling factor. Then the initial scaling factor is s _s =1, and the termination scaling factor is s _s =min{W _i /W _d , H _i /H _d }. Sliding window scanning is performed on the infrared images to be detected at each scale, and the window judgment is performed by using the trained pedestrian classifier. If the detection window is a pedestrian frame, record the position and confidence of the detection window, and express the record as {posX, posY, width, height, score}, where posX, posY are the upper left corner of the pedestrian frame, width, height is the width and height of the pedestrian frame, and score is the confidence level.

The confidence level is the error rate recorded by all the leaf nodes of the two-level decision tree.

5-2. After multi-scale sliding window scanning detection of the infrared image to be detected, the same pedestrian is detected in different scales of the infrared image to be detected. In order for the system to output a pedestrian frame that most likely corresponds to the actual pedestrian position, The non-maximum suppression method is used to fuse the pedestrian frame results of multiple different scales. The criterion for maximum value suppression is the confidence score of each pedestrian frame.

5-3. Arrange all the pedestrian frames into an array A[n] according to the confidence level from low to high, and then take out the detection window information A[n] with the highest current confidence level from the array A[n];

5-4. Determine the relationship between the pedestrian frame A[n] and the subsequent pedestrian frame A[n-1]. If the coincidence degree α of the two pedestrian frames is greater than 0.5, it is considered to be the same pedestrian frame, otherwise the pedestrian frame A [n-1] as the pedestrian frame with the highest current confidence.

The degree of coincidence Among them, area(B _n ) is the area of the nth pedestrian frame, and area(B _n ∩B _n-1 ) is the intersection of the area of the nth pedestrian frame and the area of the n-1th pedestrian frame.

5-5. Repeat steps 5-4 until all pedestrian frames are determined.

Step 6: Pedestrian frame weight determination.

6-1. After passing step 5, each detection window that is determined to be a pedestrian will have a confidence value score. If the classifier has a clearer judgment on a detection window, the corresponding score value will be larger; on the contrary, if the classifier is less clear on the judgment of the detection window, a very low score value will be obtained. Therefore, when an area is wrongly judged as a pedestrian frame, the confidence score obtained by the pedestrian frame is relatively low. In order to reduce the erroneous judgment made by the classifier when the confidence level is not enough, the present invention adopts the brightness interval template re-judgment technology. When score _i ≤τ, the pedestrian frame re-judgment is performed, and the value of the threshold τ is determined by a statistical method. If score _i >τ, then only the classifier is used for pedestrian detection, assuming that the confidence of N pedestrian frames is obtained (the default value N is 1000), then τ satisfies the formula

The score _i is the confidence value of the i-th in the array A[n] defined in step 5-3.

6-2. According to step 2 and the characteristics of infrared pedestrian imaging, the four intervals [0,t ₁ ),[t ₁ ,t ₂ ),[t ₂ ,t ₃ ),[t ₃ ,255] correspond to the background, Upper body, lower body, head. Let the width and height of the pedestrian be w×h, if the pedestrian frame correctly determines the position of the pedestrian, then divide the height h of the pedestrian frame that has been judged as a pedestrian into 4 equal parts from top to bottom, then in the first 1/4 equal part There must be brightness information belonging to the range of [t ₃ ,255], and the total number of pixels with gray values in this range is Σp ₁ , and the second 1/4 equal division must exist at [t ₁ ,t ₂ ) The brightness information of the range, and record the total number of pixels of the gray value in this interval as Σp ₂ , and there must be brightness information in the range [t ₂ ,t ₃ ) at the last two 1/4 equal parts, and record the gray The total number of pixels with degrees in this range is Σp ₃ . Set the verification rules as follows: Σp ₁ /(w×h)≥1/16, Σp ₂ /(w×h)≥1/8, Σp ₃ /(w×h)≥1/16. If Σp ₁ , Σp ₂ , and Σp ₃ satisfy these three conditions at the same time, then the pedestrian frame is judged as a pedestrian frame, otherwise the pedestrian frame is judged as a non-pedestrian frame.

Claims (2)

1. a nighttime pedestrian detection method based on infrared pedestrian brightness statistical features, is characterized in that comprising the steps: Step 1. Construct positive and negative sample data sets of infrared images; Step 2, infrared image DBHOI feature construction based on brightness statistical features; Step 3, infrared image HOG feature construction based on contour information; Step 4, utilize Adaboost to carry out classifier training; Step 5, detection window judgment and positioning; Step 6, pedestrian frame weight determination; The details of constructing the positive and negative sample data sets of infrared images described in step 1 are as follows: The method of constructing the positive sample data set is as follows: use the smallest rectangular window to extract pedestrian samples in the infrared image; assume that the height of the pedestrian is h and the width is w, so that w/h=0.41; a total of N ₁ positive samples of pedestrians are extracted; The negative sample data set construction method is as follows: Randomly extract N ₀ × 10 negative samples from N ₀ infrared images that do not contain pedestrians, that is, randomly select 10 negative samples from each infrared image; Scale all positive and negative samples to a sample image with a width of 64 pixels and a height of 128 pixels; The infrared image DBHOI feature construction based on the brightness statistical feature described in step 2 is specifically as follows: 2-1. Set the size of the sample image as w′×h′, then divide the sample image into multiple partial images of the same size; the size of each partial image is 8×8, that is, divide the sample image into equal (w′ /8)×(h′/8) partial images, and record these partial images as cells; 2-2. Determine the mapping rules according to the pedestrian components; 2-2-1. Firstly, according to N ₁ positive samples, each part of the pedestrian is intercepted, and each part of the pedestrian includes the head, upper body, and lower body; 2-2-2. Randomly intercept N ₁ background images in the negative sample; 2-2-3. Calculate the gray mean value of the head image, upper body image, lower body image and background image respectively; 2-2-4. Set the average gray values of the four images as G ₁ , G ₂ , G ₃ , and G ₄ , and the sizes increase sequentially; and determine three mapping boundaries according to the four average gray values, respectively t ₁ =(G ₁ +G ₂ )/2, t ₂ =(G ₂ +G ₃ )/2, t ₃ =(G ₃ +G ₄ )/2, divide the gray value range into four gray values according to the mapping boundary degree interval, respectively [0,t ₁ ),[t ₁ ,t ₂ ),[t ₂ ,t ₃ ),[t ₃ ,255]; 2-2-5. According to the mapping rule and using the gray value of the pixel in the infrared image as the weight, construct a brightness histogram in each cell to obtain a four-dimensional feature vector; 2-3. Adjacent 2×2 cells are normalized within the block; Taking one cell as the step size, put (w'/8)×(h'/8) cells in the sample image in order from top to bottom and from left to right to pass the adjacent four cells through L1-Sqrt method for intra-block normalization; The L1-Sqrt method is as follows: v is the brightness histogram vector, ε is a very small value, the value is 0.001; 2-4. Concatenate the features of all blocks to obtain the DBHOI features; The DBHOI is a brightness histogram of different interval sizes; The HOG feature construction of the infrared image based on the contour information described in step 3 is as follows: Use the classic sobel operator to calculate the gradient components G _x (i, j) and G _y (i, j) of each pixel in the horizontal and vertical directions, and then calculate the corresponding gradient size G (i, j) and direction D (i, j) as follows: <mrow><mi>G</mi><mrow><mo>(</mo><mi>i</mi><mo>,</mo><mi>j</mi><mo>)</mo></mrow><mo>=</mo><msqrt><mrow><msub><mi>G</mi><mi>x</mi></msub><msup><mrow><mo>(</mo><mi>i</mi><mo>,</mo><mi>j</mi><mo>)</mo></mrow><mn>2</mn></msup><mo>+</mo><msub><mi>G</mi><mi>y</mi></msub><msup><mrow><mo>(</mo><mi>i</mi><mo>,</mo><mi>j</mi><mo>)</mo></mrow><mn>2</mn></msup></mrow></msqrt></mrow> <mrow><mi>D</mi><mrow><mo>(</mo><mi>i</mi><mo>,</mo><mi>j</mi><mo>)</mo></mrow><mo>=</mo><mi>a</mi><mi>r</mi><mi>c</mi><mi>t</mi><mi>a</mi><mi>n</mi><mrow><mo>(</mo><mfrac><mrow><msub><mi>G</mi><mi>y</mi></msub><mrow><mo>(</mo><mi>i</mi><mo>,</mo><mi>j</mi><mo>)</mo></mrow></mrow><mrow><msub><mi>G</mi><mi>x</mi></msub><mrow><mo>(</mo><mi>i</mi><mo>,</mo><mi>j</mi><mo>)</mo></mrow></mrow></mfrac><mo>)</mo></mrow><mo>;</mo></mrow> The classifier training using Adaboost described in step 4 is as follows: 4-1. Scale all positive and negative samples to the same scale, then extract the DBHOI feature vector and HOG feature vector for each sample image, and mark the label of the positive sample as 1, and the label of the negative sample as -1; 4-2. Input the DBHOI-HOG feature vectors and sample labels corresponding to N ₁ positive samples and N ₀ × 10 negative samples to the Adaboost learning algorithm for training, and obtain a classifier composed of a series of weak classifiers; The determination and positioning of the detection window described in step 5 are as follows: 5-1. Determine the scaling factor according to the scale of the infrared image to be detected and the size of the detection window, and determine the number of scaling layers of the infrared image to be detected; then use the classifier obtained in step 4 to scan layer by layer with a step size of 4 pixels ; Suppose the original size of the infrared image to be detected is W _i ×H _i , where W _i represents the width of the infrared image to be detected, H _i represents the height of the infrared image to be detected, and the size of the detection window is W _d ×H _d , where W _d represents The width of the detection window, H _d represents the height of the detection window, and s _s represents the scaling factor; then the initial scaling factor is s _s =1, and the termination scaling factor is s _s =min{W _i /W _d ,H _i /H _d } ; Sliding window scanning is performed on the infrared image to be detected at each scale, and the detection window is judged by using the trained pedestrian classifier; if the detection window is a pedestrian frame, record the position and confidence of the detection window, and record the Expressed as {posX, posY, width, height, score}, where posX and posY are the upper left corner of the pedestrian frame, width and height are the width and height of the pedestrian frame, and score is the confidence level; The confidence degree is the error rate recorded by all the leaf nodes of the two-layer decision tree; 5-2. After multi-scale sliding window scanning detection of the infrared image to be detected, the same pedestrian is detected in different scales of the infrared image to be detected, and the non-maximum value suppression method is used to detect multiple pedestrian frames of different scales. The results are merged; The criterion for maximum value suppression is the confidence score of each pedestrian frame; 5-3. Arrange all the pedestrian frames into an array A[n] according to the confidence level from low to high, and then take out the detection window information A[n] with the highest current confidence level from the array A[n]; 5-4. Determine the relationship between the pedestrian frame A[n] and the subsequent pedestrian frame A[n-1]. If the coincidence degree α of the two pedestrian frames is greater than 0.5, it is considered to be the same pedestrian frame, otherwise the pedestrian frame A [n-1] as the pedestrian frame with the highest current confidence; The degree of coincidence Where area(B _n ) is the area of the nth pedestrian frame, area(B _n ∩B _n-1 ) is the intersection of the area of the nth pedestrian frame and the area of the n-1th pedestrian frame; 5-5. Repeat steps 5-4 until all pedestrian frames are determined. 2. a kind of night pedestrian detection method based on infrared pedestrian luminance statistical feature according to claim 1, it is characterized in that the pedestrian frame heavy determination described in step 6 is as follows: 6-1. After passing step 5, each detection window that is judged as a pedestrian will have a confidence value score; the pedestrian frame is re-judged using the brightness interval template; when the score _i ≤τ, the pedestrian frame is re-determined Judgment, where the value of the threshold τ is determined by a statistical method; if score _i > τ, only the classifier is used for pedestrian detection, assuming that the confidence of N pedestrian frames is obtained, then τ satisfies the formula The score _i is the i-th confidence value in the array A[n] defined in step 5-3; 6-2. According to step 2 and the characteristics of infrared pedestrian imaging, the four intervals [0,t ₁ ),[t ₁ ,t ₂ ),[t ₂ ,t ₃ ),[t ₃ ,255] correspond to the background, Upper body, lower body, head; if the width and height of the pedestrian is w×h, if the pedestrian frame correctly determines the position of the pedestrian, then the height h of the pedestrian frame that has been judged as a pedestrian is divided into 4 equal parts from top to bottom, then at There must be brightness information belonging to the range of [t ₃ ,255] at a 1/4 equal division, and the total number of pixels with gray values in this interval is ∑p ₁ , and the second 1/4 equal division must be There is brightness information in the range of [t ₁ ,t ₂ ), and the total number of pixels with gray value in this interval is ∑p ₂ , and there must be [t ₂ ,t ₃ ) at the last two 1/4 equal parts Brightness information within the range, and record the total number of pixels with gray values in this range as ∑p ₃ ; set the following verification rules: ∑p ₁ /(w×h)≥1/16, ∑p ₂ /(w×h h)≥1/8, ∑p ₃ /(w×h)≥1/16; if ∑p ₁ , ∑p ₂ , ∑p ₃ meet these three conditions at the same time, then the pedestrian frame is judged as a pedestrian frame again , otherwise the pedestrian frame is judged as a non-pedestrian frame.