CN104751136B - A kind of multi-camera video event back jump tracking method based on recognition of face - Google Patents

Abstract

The invention discloses a multi-camera video event retrospective tracking method based on face recognition, which first trains a face cascade classifier through the Adaboost method based on Haar-like features, and then uses this classifier to detect and track the events that appear in each camera Face, get the face database, frame the face area of the target pedestrian in the video, extract its LBP features, and match and identify the target face in the face database, and finally extract the target in each face according to the face recognition results The time at which the camera appears gives its walking path, thereby obtaining the backtracking result of the target. In the present invention, by applying the face recognition method in multiple camera video systems in multiple scenes, the video events are traced back, reducing the workload of manual video query and improving the query efficiency.

Description

A multi-camera video event backtracking method based on face recognition

technical field

The invention belongs to the technical field of intelligent analysis systems, and in particular relates to a multi-camera video event retrospective tracking method based on face recognition.

Background technique

Video surveillance systems are everywhere in people's lives. When an abnormal event occurs, it is usually necessary to find suspicious personnel targets from the surveillance video. The traditional method is often to check manually. Due to the large amount of video surveillance data, manual search is time-consuming, labor-intensive and inefficient, and it is easy to miss a lot of useful information. Therefore, through the processing of monitoring data by machines, the intelligent realization of backtracking has important practical application value.

For the pedestrian target in the video, although there are many kinds of features, the human face is a reliable and superior biometric feature. In recent years, with the continuous advancement of science and technology, face detection and recognition technology has also been continuously developed. As a biological feature of human beings, the face has more and more applications. For example, face recognition access management system, face recognition access control and attendance system, face recognition monitoring management, face recognition computer security protection, face recognition photo search, face recognition visitor registration, face recognition ATM intelligent video alarm system Wait. In the 2008 Olympic Games and the 2010 Shanghai World Expo, face recognition was used for identity authentication for security work. Most of these are used for a single image, or a single-camera single-scene video, and less for multi-camera multi-scene video systems. If you need to view video events in multi-camera and multi-scene videos, you need to manually check the videos of multiple cameras separately, which has the problems of time-consuming, labor-intensive, and low retrieval efficiency.

Contents of the invention

The purpose of the present invention is to provide a multi-camera video event retrospective tracking method based on face recognition to solve the problems of time-consuming, labor-intensive and low retrieval efficiency in manually viewing video events.

The technical solution of the present invention is realized in the following way: a method for backtracking and tracking of multi-camera video events based on face recognition, specifically implemented according to the following steps:

Step 1: Train the face cascade classifier through the Adaboost method based on Haar-like features;

Step 2: collect the video of all cameras in the monitoring system;

Step 3: Track the faces in all the collected camera videos, and write the tracked faces into the database as the face database. The naming rule of each face in the database is: camera number_frame number;

Step 4: Select the target pedestrian that needs to be traced back, frame the face area of the target pedestrian in the video, and match and identify the target pedestrian in the face database by extracting the LBP features of the target pedestrian and the face in the face database;

Step 5: According to the face recognition result in step 4, extract the time when the target pedestrian appeared in each camera to obtain the walking path, and thus obtain the backtracking result of the target pedestrian.

The present invention is also characterized in that,

The specific process of training face cascade classifier in step 1 is as follows:

1.1. Given the training samples, first you need to give the training sample set (x ₁ ,y ₁ ),(x ₂ ,y ₂ ),...,(x _n ,y _n ), which contains positive and negative face samples Sample, x _i represents the sample, y _i represents the positive or negative of the sample, y _i =1 represents it is a face positive sample, y _i =0 represents it is a face negative sample (non-face), n represents the total number of training samples ;

1.2, Initialize weights:

Among them, w _1,i represents the initialization weight of the i-th sample in the first iteration process, l, m are the number of positive samples and negative samples respectively, i=1,2...n;

1.3, normalized weights, normalize the weights of all samples, as shown in formula (2), q _{t, i} are the normalized weights:

Among them, w _t,i represents the weight of the i-th sample in the t-th iteration process;

1.4, train the optimal weak classifier for each Haar-like feature, and calculate its weighted error rate:

For each Haar-like feature, the optimal weak classifier h(x, f, p, θ) is trained on the basis of the sample as:

Among them _, f represents the feature, f(x) represents the feature value, p represents the direction of the inequality sign, p takes 1 or -1, θ is the threshold, and p is to make the direction of the inequality sign always be <sign;

Calculate the weighted error rate ε _f of the optimal weak classifier h(x,f,p,θ) for all samples, as shown in the following formula:

1.5. Select the optimal weak classifier with the smallest weighted error rate for all samples as the weak classifier obtained in this iteration;

1.6. Adjust the weight. The adjustment of the weight is based on whether the weak classifier obtained in step 1.5 judges each sample positively. As shown in formulas (5) and (6), it can be seen that if the positive judgment is positive, the weight is reduced. The weight of misjudgment increases, which can make the subsequent selection of weak classifiers select classifiers that pay more attention to misjudgment samples:

Among them, the value of e _i is 0 or 1, 0 when the sample is positively judged, and 1 when the sample is misjudged, ε _t represents the weighted error rate of the weak classifier for all samples obtained in the t-th iteration process ;

Repeat steps 1.3 to 1.6 for the iterative process until the specified number of iterations T is reached, T weak classifiers are obtained, and the iterative process ends;

1.7. Combine T weak classifiers into a strong classifier, and the strong classifier is shown in the following formula:

1.8. Change the number of iterations T to obtain different strong classifiers, and cascade the strong classifiers in the order of T from small to large to obtain the final cascaded classifier. The number of strong classifiers is recorded as M, and M is greater than 20.

The specific process of face tracking in step 3 is as follows:

3.1, first grayscale the video frame to be detected in the camera, then reduce the image by bilinear interpolation method to improve the detection speed, and finally perform histogram equalization to improve the detection effect;

3.2. Perform face detection and tracking on the video. The method of tracking is: for each face detected in the Nth frame, determine the possible range of its appearance in the N+1th frame according to the inter-frame constraints. The specific method is: according to The position and size of the face detected in the Nth frame is centered on the center of the face detected in the Nth frame in the N+1 frame, and the width and height of the face area detected in the Nth frame are increased Double, determine the range where the face may appear in the N+1th frame, and detect the face in the N+1th frame in this range:

If there is only one face detected within this range, it is this person;

If more than one face is detected within this range, calculate the Euclidean distance between the face detected in the Nth frame and the center of each face detected in the N+1 frame, and the face closest to the Nth frame Determine the face of this person appearing in frame N+1;

If no face is detected within this range, it is considered that the person has walked out of the video area of the camera.

The face detection method in step 3 is as follows:

a, if it is the first frame of the video image, then directly use the cascade classifier obtained in step 1 to detect;

If it is a subsequent frame of the video image, it is determined as the area where the face may appear in the frame according to the inter-frame constraints, and the first N strong classifiers of the cascade classifier obtained in step 1 are used to detect in this area, N< M; other areas except this area are detected using the cascade classifier obtained in step 1;

b. Perform skin color detection on the detected faces, and exclude non-skin color false detection results, specifically:

First convert the image from the RGB color space to the YC _b C _r color space, as shown in formulas (9), (10), and (11):

Y=0.257R+0.504G+0.098B+16 (9);

Cb=-0.148R-0.291G+0.439B+128 (10);

Cr=0.439R-0.368G-0.071B+128 (11);

Judge the Cb and Cr values of the pixels in the detected face area. If 85<Cb<130 and 132<Cr<180 are satisfied, the current pixel is a skin color point, otherwise it is a non-skin color point. In the result, if the skin color points account for more than 60% of the total pixels, it is considered as a face area, otherwise it is a non-face area and is excluded.

The specific process of step 4 is as follows:

4.1. Frame the face area of the target pedestrian in the video, and use the cascade classifier trained in step 1 to detect the target face in this area as a mark for the target pedestrian;

4.2. Extract the face of the target pedestrian and the uniform mode LBP feature of the face in the face database;

4.3. Calculate the distance between the LBP eigenvalues of the face in the face library and the face of the target pedestrian, and sort the faces in the library according to the calculated distance from small to large. According to the sorting order, select a picture with The face image closest to the face of the target pedestrian is the face image of the target pedestrian.

The specific process of step 4.2 is as follows:

4.2.1, first divide the face of the target pedestrian and the face image in the face database into m×m image blocks evenly, that is, m ² image blocks, and then judge the LBP binary code of each pixel of each image block Combination mode:

Each pixel of the image block is taken as the center, and its gray value is used as the threshold, and the gray value of the surrounding pixels is compared with it for binarization to obtain the binary code of LBP, and the surrounding pixels are compared with the central pixel for binarization The formula is as follows:

Among them, s(p) is the binarization value of the neighborhood centered on the current pixel, g _c is the gray value of the center pixel, and g _p is the gray value of the neighborhood pixel of g _c ;

The calculation formula for the number of jumps between 0 and 1 in the LBP binary code sequence is as follows:

Where s(i) represents the binarized value of the pixel P+1 neighborhood;

When the binary code combination mode of U(LBP)≤2 is a uniform mode, other binary code combination modes are classified as a non-uniform mode;

4.2.2, Find the histogram of the LBP uniform mode of each image block, and connect the m ² histograms in series as the LBP feature of the target pedestrian and the face in the face database.

The specific process of step 5 is as follows:

According to the face image selected in step 4.3, its frame number can be obtained through its naming rule in step 3, and further according to the starting time and frame rate of the camera video, the time when the face appears in the video can be obtained, Finally, the face video frames selected in each camera are marked in the order of appearance time, and the backtracking result of the target pedestrian is obtained.

The beneficial effect of the present invention is that the present invention intelligently obtains the time when the target pedestrian appears in each camera video through face recognition, so that the target pedestrian can be efficiently traced back and tracked according to the installation position of the camera. The depiction of the action track can quickly find its whereabouts when an abnormal event occurs, without manually checking the video all the time, reducing the workload of manual search.

Description of drawings

Fig. 1 is a flow chart of a method for backtracking and tracking multi-camera video events based on face recognition in the present invention;

Fig. 2 is the process of the training cascade classifier of the present invention;

Fig. 3 is a partial face image of a person in the face database obtained by tracking in the present invention;

Fig. 4 is the image of the face area of the target pedestrian framed in the present invention;

Fig. 5 is the image of the target face detected in the present invention;

Fig. 6 is the image of the target pedestrian in each camera obtained by backtracking using the backtracking method of the present invention;

Fig. 7 is a schematic diagram of the final backtracking result of the target pedestrian in each camera under multi-camera obtained backtracking by using the backtracking method of the present invention.

Detailed ways

The present invention will be described in detail below with reference to the drawings and specific embodiments.

Referring to Fig. 1, a method for retrospectively tracking multi-camera video events based on face recognition of the present invention is specifically implemented according to the following steps:

Step 1: Train the face cascade classifier through the Adaboost method based on Haar-like features. The specific process is as follows:

In face detection, each weak classifier that composes the strong classifier trained by the Adaboost algorithm corresponds to a Haar-like feature, which is the optimal weak classifier based on this feature, representing the grayscale of a certain part of the face distribution characteristics. Since not every face feature described by Haar-like features is important, it can be used as a feature to distinguish between a face and a non-face, so the Adaboost algorithm needs to be iteratively selected from these very large features. The features of face and non-face classification are used as the selected weak classifiers, and each weak classifier is assigned a corresponding weight according to its classification ability, thus forming the final strong classifier.

Referring to Figure 2, the specific process of training a strong classifier is as follows:

1.1. Given the training samples, first you need to give the training sample set (x ₁ ,y ₁ ),(x ₂ ,y ₂ ),...,(x _n ,y _n ), which contains positive and negative face samples Sample, x _i represents the sample, y _i represents the positive or negative of the sample, y _i =1 represents it is a face positive sample, y _i =0 represents it is a face negative sample (non-face), n represents the total number of training samples ;

1.2, Initialize weights:

In the formula, w _1,i represents the initialization weight of the i-th sample in the first iteration process, l, m are the number of positive samples and negative samples respectively, where i=1,2...n;

1.3, normalized weights, normalize the weights of all samples, as shown in formula (2), q _{t, i} are the normalized weights:

Among them, w _t,i represents the weight of the i-th sample in the t-th iteration process;

1.4, train the optimal weak classifier for each Haar-like feature, and calculate its weighted error rate:

For each Haar-like feature, the optimal weak classifier h(x, f, p, θ) is trained on the basis of the sample as:

Among them, f represents the feature, f(x) represents the feature value, θ is the threshold, p represents the direction of the inequality sign, p takes 1 or -1, and p is to make the direction of the inequality sign always be < sign;

Calculate the weighted error rate ε _f of the optimal weak classifier h(x,f,p,θ) for all samples, as shown in the following formula:

1.5. Select the optimal weak classifier with the smallest weighted error rate for all samples as the weak classifier obtained in this iteration;

1.6. Adjust the weight. The adjustment of the weight is based on whether the weak classifier obtained in step 1.5 judges each sample positively. As shown in formulas (5) and (6), it can be seen that if the positive judgment is positive, the weight is reduced. The weight of misjudgment increases, which can make the subsequent selection of weak classifiers select classifiers that pay more attention to misjudgment samples:

Among them, the value of e _i is 0 or 1, 0 when the sample is positively judged, and 1 when the sample is misjudged, ε _t represents the weighted error rate of the weak classifier for all samples obtained in the t-th iteration process ;

Repeat steps 1.3 to 1.6 for the iterative process until the specified number of iterations T is reached, T weak classifiers are obtained, and the iterative process ends;

1.7. Combine T weak classifiers into a strong classifier. The strong classifier is composed of weights based on the classification ability of the weak classifiers, that is, each weak classifier votes to determine the judgment result, and only the component voted by each weak classifier Not the same, the obtained strong classifier is as follows:

It is equivalent to letting all weak classifiers vote, and then weighting and summing the voting results according to the error rate of the weak classifiers, and comparing the weighted and summed results of the votes with the average voting results to obtain the final result;

1.8. Change the number of iterations T to obtain different strong classifiers, and cascade the strong classifiers in the order of T from small to large to obtain the final cascaded classifier. The number of strong classifiers is recorded as M, and M is greater than 20.

Step 2: collect the video of all cameras in the monitoring system;

Step 3: Track the faces in all the collected camera videos, get multiple faces of each person and use them as a face library, as shown in Figure 3, the specific process is as follows:

3.1, first grayscale the video frame to be detected in the camera, then reduce the image by bilinear interpolation method to improve the detection speed, and finally perform histogram equalization to improve the detection effect;

3.2. Perform face detection and tracking on the video. The method of tracking is: for each face detected in the Nth frame, determine the possible range of its appearance in the N+1th frame according to the inter-frame constraints. The specific method is: according to The position and size of the face detected in the Nth frame is centered on the center of the face detected in the Nth frame in the N+1 frame, and the width and height of the face area detected in the Nth frame are increased Double, determine the range where the face may appear in the N+1th frame, and detect the face in the N+1th frame in this range:

If there is only one face detected within this range, it is this person;

If more than one face is detected within this range, calculate the Euclidean distance between the face detected in the Nth frame and the center of each face detected in the N+1 frame, and the face closest to the Nth frame Determine the face of this person appearing in frame N+1;

If no face is detected within this range, it is considered that the person has walked out of the video area of the camera;

The face detection method is as follows:

a, if it is the first frame of the video image, then directly use the cascade classifier obtained in step 1 to perform face detection;

If it is a subsequent frame of the video image, it is determined according to the inter-frame constraints as the area where the face to be followed in the frame may appear, and the first N strong classifiers of the cascade classifier obtained in step 1 are used for detection in this area , N<M; other areas except this area are detected using the cascade classifier obtained in step 1;

b. Perform skin color detection on the detected faces, and exclude non-skin color false detection results, specifically:

First convert the image from the RGB color space to the YC _b C _r color space, as shown in formulas (9), (10), and (11):

Y=0.257R+0.504G+0.098B+16 (9);

Cb=-0.148R-0.291G+0.439B+128 (10);

Cr=0.439R-0.368G-0.071B+128 (11);

Judge the Cb and Cr values of the pixels in the detected face area. If 85<Cb<130 and 132<Cr<180 are satisfied, the current pixel is a skin color point, otherwise it is a non-skin color point. In the result, if the skin color points account for more than 60% of the total pixels, it is considered as a face area, otherwise it is a non-face area and is excluded.

3.3. Write the face obtained after tracking into the database as a face library. The naming rule of each face in the database is: camera number_frame number, where the camera number indicates which camera the face belongs to and is detected in the video. The frame number indicates which frame of the camera video the face image belongs to.

Step 4, frame the face area of the target pedestrian in the video, and perform matching and recognition on the target face in the face database by extracting LBP features; the specific process is as follows:

4.1, frame the face area of the target pedestrian in the video, as shown in Figure 4, use the cascade classifier trained in step 1 in this area, and detect the target face in this area as a mark for the target pedestrian , as shown in Figure 5;

4.2. Extract the face of the target pedestrian and the uniform mode LBP feature of the face in the face database:

4.2.1 First divide the face of the target pedestrian and the face image in the face database into 4×4 image blocks evenly, that is, 16 image blocks, and then judge the LBP binary code combination mode of each pixel of each image block :

LBP is an operator that describes local texture features. The current pixel is taken as the center, and its gray value is used as a threshold. The gray value of the surrounding pixels is compared with it for binarization, and the binary code of LBP is obtained. The surrounding pixels The formula for comparing binarization between points and central pixels is as follows:

Among them, s(p) is the binarization value of the neighborhood centered on the current pixel, g _c is the gray value of the center point, and g _p is the gray value of the neighborhood pixel of g _c ;

How many values does the binary code of LBP have, that is, how many different binary code combination modes there are. However, after research, it is found that some patterns represent most of the texture patterns, and the probability of appearing in the image is as high as 90%, while the probability of some other patterns appearing in the image is very low, so it can be considered that some patterns are the basic texture of the image. Attributes, these patterns are uniform patterns (Uniform Pattern), that is, patterns in which the number of jumps between 0 and 1 in the LBP binary code sequence does not exceed two times, and the calculation formula for the number of jumps is as follows:

Where s(i) represents the binarized value of the pixel P+1 neighborhood;

When the binary code combination mode of U(LBP)≤2 is a uniform mode, other binary code combination modes are classified as a non-uniform mode;

4.2.2, Find the histogram of the LBP uniform mode of each image block, and connect these 16 histograms in series, as the LBP feature of the target pedestrian and the face in the face database;

4.3. Through the Euclidean distance, calculate the distance between the face in the face library and the LBP feature value between the face of the target pedestrian, and sort the faces in the library according to the calculated distance from small to large. According to the sorting order, select A face image that is closest to the face of the target pedestrian is generated, which is the face image of the target pedestrian.

Step 5, according to the face recognition result, extract the time when the target pedestrian appears in each camera to get his walking path, and thus obtain the backtracking tracking result of the target. The specific process is as follows:

According to the face image selected in step 4.3, its frame number can be obtained through its naming rule in step 3, and further according to the starting time and frame rate of the camera video, the time when the face appears in the video can be obtained, Finally, the face video frames selected in each camera are marked in the order of appearance time, and the backtracking result of the target pedestrian can be obtained.

Figure 7 shows the final backtracking result diagram. From Figure 7, it can be seen that the successive positions of the target pedestrians are the corresponding positions of camera 5, camera 4, camera 3, camera 1, and camera 2. Therefore, using the method of the present invention can Quickly and accurately realize the backtracking of the target, which greatly improves the retrieval efficiency of the target pedestrian in the video event.

Claims (5)

1. A method for backtracking and tracking of multi-camera video events based on face recognition, characterized in that, it is specifically implemented according to the following steps: Step 1: Train the face cascade classifier through the Adaboost method based on Haar-like features; Step 2: collect the video of all cameras in the monitoring system; Step 3: Track the faces in all the captured camera videos, and write the tracked faces into the database as a face database. The naming rule of each face in the database is: camera number_frame number: The specific process of face tracking is as follows: 3.1, first grayscale the video frame to be detected in the camera, then reduce the image by bilinear interpolation method to improve the detection speed, and finally perform histogram equalization to improve the detection effect; 3.2. Perform face detection and tracking on the video. The following method is: for each face detected in the Nth frame, determine the possible range of the N+1th frame according to the inter-frame constraints. The specific method is: according to The position and size of the face detected in the Nth frame is centered on the center of the face detected in the Nth frame in the N+1 frame, and the width and height of the face area detected in the Nth frame are increased Double, determine the range where the face may appear in the N+1th frame, and detect the face in the N+1th frame in this range: If there is only one face detected within this range, it is this person; If more than one face is detected within this range, calculate the Euclidean distance between the face detected in the Nth frame and the center of each face detected in the N+1 frame, and the face closest to the Nth frame Determine the face of this person appearing in frame N+1; If no face is detected within this range, it is considered that the person has walked out of the video area of the camera; The face detection method in step 3 is as follows: a, if it is the first frame of the video image, then directly use the cascade classifier obtained in step 1 to detect; If it is a subsequent frame of the video image, it is determined as the area where the face may appear in the frame according to the inter-frame constraints, and the first N strong classifiers of the cascade classifier obtained in step 1 are used to detect in this area, N< M; other areas except this area are detected using the cascade classifier obtained in step 1; b. Perform skin color detection on the detected faces, and exclude non-skin color false detection results, specifically: First convert the image from the RGB color space to the YC _b C _r color space, as shown in formulas (9), (10), and (11): Y=0.257R+0.504G+0.098B+16 (9); Cb=-0.148R-0.291G+0.439B+128 (10); Cr=0.439R-0.368G-0.071B+128 (11); Judge the Cb and Cr values of the pixels in the detected face area. If 85<Cb<130 and 132<Cr<180 are satisfied, the current pixel is a skin color point, otherwise it is a non-skin color point. In the result, if the skin color points account for more than 60% of the total pixels, it is considered as a face area, otherwise it is a non-face area and is excluded; Step 4: Select the target pedestrian that needs to be traced back, frame the face area of the target pedestrian in the video, and match and identify the target pedestrian in the face database by extracting the LBP features of the target pedestrian and the face in the face database; Step 5: According to the face recognition result in step 4, extract the time when the target pedestrian appeared in each camera to obtain the walking path, and thus obtain the backtracking result of the target pedestrian. 2. a kind of multi-camera video event retrospective tracking method based on face recognition according to claim 1, is characterized in that, the concrete process of training face cascade classifier in step 1 is as follows: 1.1. Given the training samples, first you need to give the training sample set (x ₁ ,y ₁ ),(x ₂ ,y ₂ ),...,(x _n ,y _n ), which contains positive and negative face samples Sample, x _i represents the sample, y _i represents the positive or negative of the sample, y _i =1 represents it is a face positive sample, y _i =0 represents it is a face negative sample (non-face), n represents the total number of training samples ; 1.2, Initialize weights: <mrow><msub><mi>w</mi><mrow><mn>1</mn><mo>,</mo><mi>i</mi></mrow></msub>< mfenced open = "{" close = ""><mtable><mtr><mtd><mfrac><mn>1</mn><mrow><mn>2</mn><mi>l</mi></mrow></mfrac></mtd><mtd><mrow><msub><mi>y</mi><mi>i</mi></msub><mo>=</mo><mn>1</mn></mrow></mtd></mtr><mtr><mtd><mfrac><mn>1</mn><mrow><mn>2</mn><mi>m</mi></mrow></mfrac></mtd><mtd><mrow><msub><mi>y</mi><mi>i</mi></msub><mo>=</mo><mn>0</mn></mrow></mtd></mtr></mtable></mfenced><mo>-</mo><mo>-</mo><mo>-</mo><mrow><mo>(</mo><mn>1</mn><mo>)</mo></mrow><mo>;</mo></mrow> Among them, w _1,i represents the initialization weight of the i-th sample in the first iteration process, l, m are the number of positive samples and negative samples respectively, i=1,2...n; 1.3, normalized weights, normalize the weights of all samples, as shown in formula (2), q _{t, i} are the normalized weights: <mrow><msub><mi>q</mi><mrow><mi>t</mi><mo>,</mo><mi>i</mi></mrow></msub><mo>=</mo><mfrac><msub><mi>w</mi><mrow><mi>t</mi><mo>,</mo><mi>i</mi></mrow></msub><mrow><msubsup><mo>&amp;Sigma;</mo><mrow><mi>i</mi><mo>=</mo><mn>1</mn></mrow><mi>n</mi></msubsup><msub><mi>w</mi><mrow><mi>t</mi><mo>,</mo><mi>i</mi></mrow></msub></mrow></mfrac><mo>-</mo><mo>-</mo><mo>-</mo><mrow><mo>(</mo><mn>2</mn><mo>)</mo></mrow><mo>;</mo></mrow> Among them, w _t,i represents the weight of the i-th sample in the t-th iteration process; 1.4, train the optimal weak classifier for each Haar-like feature, and calculate its weighted error rate: For each Haar-like feature, the optimal weak classifier h(x, f, p, θ) is trained on the basis of the sample as: Among them, f represents the feature, f(x) represents the feature value, p represents the direction of the inequality sign, p takes 1 or -1, θ is the threshold, and p is to make the direction of the inequality sign always be < sign; Calculate the weighted error rate ε _f of the optimal weak classifier h(x,f,p,θ) for all samples, as shown in the following formula: <mrow><msub><mi>&amp;epsiv;</mi><mi>f</mi></msub><mo>=</mo><munderover><mo>&amp;Sigma;</mo><mrow><mi>i</mi><mo>=</mo><mn>1</mn></mrow><mi>n</mi></munderover><msub><mi>q</mi><mrow><mi>t</mi><mo>,</mo><mi>i</mi></mrow></msub><mo>|</mo><mi>h</mi><mrow><mo>(</mo><msub><mi>x</mi><mi>i</mi></msub><mo>,</mo><mi>f</mi><mo>,</mo><mi>p</mi><mo>,</mo><mi>&amp;theta;</mi><mo>)</mo></mrow><mo>-</mo><msub><mi>y</mi><mi>i</mi></msub><mo>|</mo><mo>-</mo><mo>-</mo><mo>-</mo><mrow><mo>(</mo><mn>4</mn><mo>)</mo></mrow><mo>;</mo></mrow> 1.5. Select the optimal weak classifier with the smallest weighted error rate for all samples as the weak classifier obtained in this iteration; 1.6. Adjust the weight. The adjustment of the weight is based on whether the weak classifier obtained in step 1.5 judges each sample positively. As shown in formulas (5) and (6), it can be seen that if the positive judgment is positive, the weight is reduced. The weight of misjudgment increases, which can make the subsequent selection of weak classifiers select classifiers that pay more attention to misjudgment samples: <mrow><msub><mi>w</mi><mrow><mi>t</mi><mo>+</mo><mn>1</mn><mo>,</mo><mi>i</mi></mrow></msub><mo>=</mo><msub><mi>w</mi><mrow><mi>t</mi><mo>,</mo><mi>i</mi></mrow></msub><msubsup><mi>&amp;beta;</mi><mi>t</mi><mrow><mn>1</mn><mo>-</mo><msub><mi>e</mi><mi>i</mi></msub></mrow></msubsup><mo>-</mo><mo>-</mo><mo>-</mo><mrow><mo>(</mo><mn>5</mn><mo>)</mo></mrow><mo>;</mo></mrow> <mrow><msub><mi>&amp;beta;</mi><mi>t</mi></msub><mo>=</mo><mfrac><msub><mi>&amp;epsiv;</mi><mi>t</mi></msub><mrow><mn>1</mn><mo>-</mo><msub><mi>&amp;epsiv;</mi><mi>t</mi></msub></mrow></mfrac><mo>-</mo><mo>-</mo><mo>-</mo><mrow><mo>(</mo><mn>6</mn><mo>)</mo></mrow><mo>;</mo></mrow> Among them, the value of e _i is 0 or 1, 0 when the sample is positively judged, and 1 when the sample is misjudged, ε _t represents the weighted error rate of the weak classifier for all samples obtained in the t-th iteration process ; Repeat steps 1.3 to 1.6 for the iterative process until the specified number of iterations T is reached, T weak classifiers are obtained, and the iterative process ends; 1.7. Combine T weak classifiers into a strong classifier, and the strong classifier is shown in the following formula: <mrow><msub><mi>&amp;alpha;</mi><mi>t</mi></msub><mo>=</mo><mi>l</mi><mi>o</mi><mi>g</mi><mfrac><mn>1</mn><msub><mi>&amp;beta;</mi><mi>t</mi></msub></mi>mfrac><mo>-</mo><mo>-</mo><mo>-</mo><mrow><mo>(</mo><mn>8</mn><mo>)</mo></mrow><mo>;</mo></mrow> 1.8. Change the number of iterations T to obtain different strong classifiers, and cascade the strong classifiers in the order of T from small to large to obtain the final cascaded classifier. The number of strong classifiers is recorded as M, and M is greater than 20. 3. a kind of face recognition-based multi-camera video event retrospective tracking method according to claim 1, is characterized in that, the specific process of step 4 is as follows: 4.1. Frame the face area of the target pedestrian in the video, and use the cascade classifier trained in step 1 to detect the target face in this area as a mark for the target pedestrian; 4.2. Extract the face of the target pedestrian and the uniform mode LBP feature of the face in the face database; 4.3. Calculate the distance between the LBP eigenvalues of the face in the face library and the face of the target pedestrian, and sort the faces in the library according to the calculated distance from small to large. According to the sorting order, select a picture with The face image closest to the face of the target pedestrian is the face image of the target pedestrian. 4. a kind of face recognition-based multi-camera video event retrospective tracking method according to claim 3, is characterized in that, the specific process of step 4.2 is as follows: 4.2.1, first divide the face of the target pedestrian and the face image in the face database into m×m image blocks evenly, that is, m ² image blocks, and then judge the LBP binary code of each pixel of each image block Combination mode: Each pixel of the image block is taken as the center, and its gray value is used as the threshold, and the gray value of the surrounding pixels is compared with it for binarization to obtain the binary code of LBP, and the surrounding pixels are compared with the central pixel for binarization The formula is as follows: <mrow><mi>s</mi><mrow><mo>(</mo><mi>p</mi><mo>)</mo></mrow><mo>=</mo><mfenced open = "{" close = ""><mtable><mtr><mtd><mrow><mn>1</mn><mo>,</mo></mrow></mtd><mtd><mrow><msub><mi>g</mi><mi>p</mi></msub><mo>-</mo><msub><mi>g</mi><mi>c</mi></msub><mo>&amp;GreaterEqual;</mo><mn>0</mn></mrow></mtd></mtr><mtr><mtd><mrow><mn>0</mn><mo>,</mo></mrow></mtd><mtd><mrow><msub><mi>g</mi><mi>p</mi></msub><mo>-</mo><msub><mi>g</mi><mi>c</mi></msub><mo>&lt;</mo><mn>0</mn></mrow></mtd></mtr></mtable></mfenced><mo>-</mo><mo>-</mo><mo>-</mo><mrow><mo>(</mo><mn>12</mn><mo>)</mo></mrow><mo>;</mo></mrow> Among them, s(p) is the binarization value of the neighborhood centered on the current pixel, g _c is the gray value of the center pixel, and g _p is the gray value of the neighborhood pixel of g _c ; The calculation formula for the number of jumps between 0 and 1 in the LBP binary code sequence is as follows: <mrow><mi>U</mi><mrow><mo>(</mo><mi>L</mi><mi>B</mi><mi>P</mi><mo>)</mo></mrow><mo>=</mo><munderover><mo>&amp;Sigma;</mo><mrow><mi>i</mi><mo>=</mo><mn>1</mn></mrow><mi>P</mi></munderover><mo>|</mo><mi>s</mi><mrow><mo>(</mo><mi>i</mi><mo>)</mo></mrow><mo>-</mo><mi>s</mi><mrow><mo>(</mo><mi>i</mi><mo>-</mo><mn>1</mn><mo>)</mo></mrow><mo>|</mo><mo>-</mo><mo>-</mo><mo>-</mo><mrow><mo>(</mo><mn>13</mn><mo>)</mo></mrow><mo>;</mo></mrow> Where s(i) represents the binarized value of the pixel P+1 neighborhood; When the binary code combination mode of U(LBP)≤2 is a uniform mode, other binary code combination modes are classified as a non-uniform mode; 4.2.2, Find the histogram of the LBP uniform mode of each image block, and connect the m ² histograms in series as the LBP feature of the target pedestrian and the face in the face database. 5. a kind of face recognition-based multi-camera video event retrospective tracking method according to claim 1, is characterized in that, the specific process of step 5 is as follows: According to the face image selected in step 4.3, its frame number can be obtained through its naming rule in step 3, and further according to the starting time and frame rate of the camera video, the time when the face appears in the video can be obtained, Finally, the face video frames selected in each camera are marked in the order of appearance time, and the backtracking result of the target pedestrian is obtained.