CN104463253B - Passageway for fire apparatus safety detection method based on adaptive background study - Google Patents

Abstract

The invention discloses a safety detection method for fire exits based on self-adaptive background learning. Firstly, the first N frame images of fire exit monitoring video streams are used as training samples to obtain a mixed Gaussian model, and then the background image is obtained according to the mixed Gaussian model. The image performs background clipping method on the detection image to extract the foreground area of the detection image, and judge whether the foreground area is the real foreground or the false foreground caused by other interferences through the foreground image and the HOG feature vector of the image in the foreground area of the detection image, so as to judge the detection Whether there is an abnormality in the image, if there is an abnormality in consecutive M frames, there is a safety problem in the fire exit. The invention realizes the accurate detection of whether the fire exit is safe or not by establishing a mixed Gaussian model of the background, foreground target detection and self-adaptive background update strategy based on HOG features.

Description

Fire Exit Safety Detection Method Based on Adaptive Background Learning

technical field

The invention belongs to the technical field of computer vision, and more specifically relates to a safety detection method for fire exits based on self-adaptive background learning.

Background technique

Fire exits refer to the passages used by firefighters to rescue and evacuate trapped people when various dangers occur. When an emergency such as a fire occurs, firefighters, fire vehicles and firefighting equipment must enter from the fire exit. Fire exit safety detection mainly includes two aspects: (1) fire exit blockage detection, (2) normally closed fire door abnormal opening detection. If the fire exit is occupied by motor vehicles or other obstructions, it may cause firefighters, firefighting vehicles and firefighting equipment to fail to arrive at the scene in time and delay the rescue of the fire, causing huge economic losses and casualties. The fire door is a closed door structure with a rebound function. Its function is to meet the fire resistance stability, integrity and heat insulation within a certain period of time. Special features for high temperature. Only when it is in the closed state can it effectively block the invasion of thick smoke and fire after a fire occurs, and win time for personnel evacuation and fire rescue. Therefore, it is particularly important to detect the blockage of fire exits and the abnormal opening of normally closed fire doors.

The traditional safety inspection of fire exits mainly relies on manual inspection, and special staff are appointed to check whether the fire exits are blocked and whether the normally closed fire doors are opened abnormally. This method is simple and easy, does not require any equipment, and is low in cost. The disadvantage of the first is that it cannot find the potential safety hazards in the fire escape in time; the second is that it relies heavily on the staff and is highly subjective. Another method relies on collecting video and then connecting it to the monitoring room for special staff to check. Although this method is more convenient than the first method, it reduces the burden on the staff and can monitor in a centralized manner, but it is dependent on personnel It is still very large, and it cannot achieve the purpose of real-time and automatic detection.

Contents of the invention

The purpose of the present invention is to overcome the deficiencies in the prior art, propose a fire escape safety detection method based on adaptive background learning, by setting up a mixed Gaussian model of the background, foreground target detection and adaptive background update strategy based on HOG features, to achieve Accurate detection of whether the fire escape is safe.

In order to achieve the above-mentioned purpose of the invention, the fire exit safety detection method based on adaptive background learning of the present invention comprises the following steps:

S1: Take the first N frames of images of the fire exit monitoring video stream as training samples, and train to obtain a background mixed Gaussian model mixed by three Gaussian models. The specific steps of training include:

S1.1: Initialize the mixed Gaussian model, where Gaussian model 1 is initialized according to the first frame image, the expression is:

in, Indicates the mean value of pixel j after the training of the first frame image in Gaussian model 1, the value range of j is j=1,2,...,M, M is the number of pixels in the image, x _j,1 represents the first frame image The pixel value of pixel point j in Indicates the variance of pixel j after the first frame of image training in Gaussian model 1, σ _init is the preset initial variance, Indicates the weight of pixel j after the training of the first frame image in Gaussian model 1, ω _init is the preset initial weight;

The expressions for the initialization of Gaussian models 2 and 3 are:

respectively represent the mean value of pixel j after the training of the first frame image in Gaussian model 2 and 3, Indicates the variance of pixel j after the training of the first frame image in Gaussian model 2 and 3, Represents the weight of pixel j after the training of the first frame image in Gaussian models 2 and 3;

S1.2: Use the 2nd, 3rd,...N frame images to train the mixed Gaussian model in sequence to obtain the mixed Gaussian model. The training method for the t+1th frame image is:

The three Gaussian models of the pixel point j obtained through the previous t frames of images, according to The values of i=1, 2, 3 are arranged in descending order, and each Gaussian model is matched in sequence by using the pixel value x _{j, t+1} of pixel j in the t+1th frame image. The matching method is: if Where δ is a constant greater than 0, it is considered to match, otherwise it does not match;

Once a Gaussian model is successfully matched, the Gaussian model is updated using the pixel value x _j,t+1 , and other Gaussian models remain unchanged. The update formula is:

Among them, α is the preset model learning rate, and the value range is 0<α<1, respectively represent the weight of the i-th Gaussian model in the mixed Gaussian model of pixel j after image training in the t-th frame and the t+1-th frame, respectively represent the mean value of the i-th Gaussian model in the mixed Gaussian model of the pixel j after image training in the t-th frame and the t+1-th frame, respectively represent the variance of the i-th Gaussian model in the mixed Gaussian model of the pixel j after image training in the t-th frame and the t+1-th frame;

If the pixel value x _{j, t+1} does not match the three Gaussian models, the last Gaussian model is replaced and updated according to the pixel value x _{j, t+1} , and the update formula is:

After the update is complete, the weights of the three Gaussian models To normalize, let Note that the mean value of the trained mixed Gaussian model is Variance is Weight is

S2: For the detected image in the video stream of the fire exit monitoring, the background subtraction method is used to obtain the foreground binary image of the frame image according to the mixed Gaussian model. The specific method is:

S2.1: Press the three Gaussian models of pixel j into Arrange in descending order, take the first B _j Gaussian models as the background distribution, the calculation formula of B _j is:

T represents a preset threshold;

S2.2: Match the pixel value x′j of pixel _j in the detection image to B _j Gaussian models, if any Gaussian model is successfully matched, then determine that pixel j is a background pixel, and match the corresponding pixel of the binary image The value of the point is set to 0, otherwise it is determined that the pixel point j is a foreground pixel, and the value of the corresponding pixel point of the binary image is set to 1;

S3: Use the HOG feature to judge the authenticity of the foreground block in the foreground binary image, the specific method is:

S3.1: Perform connected domain analysis on the foreground binary image obtained in step S2, classify the connected areas into one category and extract the position information of the four boundaries, and obtain the circumscribed rectangle coordinates of each connected area;

S3.2: Use the pixel mean value of the Gaussian model with the largest weight among the three Gaussian models for each pixel as the pixel value of the background pixel to form a background image;

S3.3: For each connected region, extract the corresponding gray image block from the gray image of the detection image and the background image according to the coordinates of its circumscribed rectangle, respectively extract the HOG feature vectors of the two gray image blocks, and calculate The Euclidean distance D of the two feature vectors, if D is greater than the preset threshold T _D , it is determined that the connected area is a true foreground block, otherwise it is a false foreground block;

S4: If the result analyzed in step S3 has no connected domains or each connected domain is a pseudo foreground block, then there is no abnormality in the frame detection image, if at least one connected domain is a real foreground block, then it is judged that the frame has Abnormal situation; if there are abnormal situations in consecutive M frames, and M is the preset number of frames, then there is a safety problem in the fire exit, otherwise the fire exit is safe.

The safety detection method of the fire exit based on self-adaptive background learning of the present invention first uses the first N frames of images of the fire exit monitoring video stream as training samples to obtain a mixed Gaussian model, then obtains the background image according to the mixed Gaussian model, and uses the background image to detect the image The background clipping method is used to extract the foreground area of the detected image, and the HOG feature vector of the foreground image and the image in the foreground area of the detected image is used to judge whether the foreground area is a real foreground or a false foreground caused by other interference, thereby judging whether there is Abnormal conditions, if there are abnormal conditions in consecutive M frames, there is a safety problem in the fire exit. The present invention has the following beneficial effects:

(1) The present invention self-adaptively builds the background mixed Gaussian model of monitoring scene by training samples, the background image that obtains is more in line with the actual situation, and the detection result based on this background image is more accurate;

(2) Use the HOG feature vector to judge the authenticity of the foreground target, which can effectively eliminate the interference caused by illumination changes;

(3) The mixture Gaussian model of the background can also be updated online in real time by detecting images, and the strategy of only updating the background is adopted during the update, which not only updates the scene necessary, but also avoids the excessive time for the foreground target to appear Long bugs that fade into the background.

Description of drawings

Fig. 1 is the flow chart of the specific implementation of the fire exit safety detection method based on self-adaptive background learning of the present invention;

Fig. 2 is the flowchart of the mixed Gaussian model training of background in Fig. 1;

Fig. 3 is a flow chart of foreground binary image generation in Fig. 1;

Fig. 4 is the flowchart of authenticity judgment of the foreground block in Fig. 1;

Figure 5 is a normal detection image of scene 1;

Fig. 6 is the background image obtained according to the background mixed Gaussian model at the time shown in Fig. 5;

Fig. 7 is the foreground binary image that scene 1 obtains;

Fig. 8 is the detection image of scene 2 illumination mutation;

Fig. 9 is the background image obtained by the time background mixture model shown in Fig. 8;

Fig. 10 is the foreground binary image obtained in scene 2;

Fig. 11 is the grayscale image of the detection image and the background image extracted after the connected domain analysis of scene 2, wherein Fig. 11 (a) is the grayscale image of the detection image, and Fig. 11 (b) is the grayscale image of the background image;

Fig. 12 is a histogram of the HOG feature vector of scene 2;

Figure 13 is the detection image of channel blockage in scene 3;

Fig. 14 is the background image obtained by the time background mixture model shown in Fig. 13;

Fig. 15 is the foreground binary image obtained in scene 2;

Fig. 16 is the grayscale image of the detected image and the background image extracted after the connected domain analysis of scene 3, wherein Fig. 16 (a) is the grayscale image of the detected image, and Fig. 16 (b) is the grayscale image of the background image;

Fig. 17 is a histogram of the HOG feature vector of scene 3;

Fig. 18 is a scene schematic diagram of a sudden illumination change scene, wherein Fig. 18(a) is a detection image, Fig. 18(b) is a background image, and Fig. 18(c) is a detection result diagram;

Fig. 19 is a scene schematic diagram of fire door opening scene 1, wherein Fig. 19(a) is a detection image, Fig. 19(b) is a background image, and Fig. 19(c) is a detection result diagram;

Fig. 20 is a scene schematic diagram of fire door opening scene 2, wherein Fig. 20 (a) is a detection image, Fig. 20 (b) is a background image, and Fig. 20 (c) is a detection result diagram;

Fig. 21 is a scene schematic diagram of fire exit blocking scene 1, wherein Fig. 21(a) is a detection image, Fig. 21(b) is a background image, and Fig. 21(c) is a detection result diagram;

Fig. 22 is a schematic diagram of the scene 2 of fire exit blocking scene, wherein Fig. 22(a) is a detection image, Fig. 22(b) is a background image, and Fig. 22(c) is a detection result diagram;

Fig. 23 is a schematic diagram of the outdoor fire exit scene 2, wherein Fig. 23(a) is a detection image, Fig. 23(b) is a background image, and Fig. 23(c) is a detection result diagram.

Detailed ways

Specific embodiments of the present invention will be described below in conjunction with the accompanying drawings, so that those skilled in the art can better understand the present invention. It should be noted that in the following description, when detailed descriptions of known functions and designs may dilute the main content of the present invention, these descriptions will be omitted here.

Example

FIG. 1 is a flow chart of a specific embodiment of the fire exit safety detection method based on adaptive background learning in the present invention. As shown in Figure 1, the fire escape safety detection method based on self-adaptive background learning of the present invention comprises the following steps:

S101: Gaussian mixture model training for the background:

The first N frames of images of the fire exit monitoring video stream are used as training samples, and the mixed Gaussian model of each pixel of the background mixed by three Gaussian models is obtained through training.

Fig. 2 is a flowchart of the training of the mixed Gaussian model in Fig. 1 . As shown in Figure 2, the specific steps of the mixed Gaussian model training include:

S201: Initialize the mixed Gaussian model:

First, the mixture Gaussian model is initialized. The background model of each pixel in the image in the present invention is formed by mixing three Gaussian models. Among them, the Gaussian model 1 is initialized according to the t=1th frame image, and the expression is:

in, Indicates the mean value of pixel j after the training of the first frame image in Gaussian model 1, the value range of j is j=1,2,...,M, M is the number of pixels in the image, x _j,1 represents the first frame image The pixel value of pixel point j in Indicates the variance of pixel j after the first frame of image training in Gaussian model 1, σ _init is the preset initial variance, Indicates the weight of pixel j after the first frame of image training in Gaussian model 1, ω _init is the preset initial weight. In this embodiment, the initial variance σ _init is set to 30, and the initial weight ω _init is set to 0.05.

The expressions for the initialization of Gaussian models 2 and 3 are:

respectively represent the mean value of pixel j after the training of the first frame image in Gaussian model 2 and 3, Indicates the variance of pixel j after the training of the first frame image in Gaussian model 2 and 3, Indicates the weight of pixel j after the training of the first frame image in Gaussian models 2 and 3.

S202: Gaussian model sorting:

The three Gaussian models of the pixel point j obtained through the previous t frames of images, according to The values are sorted in descending order.

S203: set i=1.

S204: Match the Gaussian model i using the t+1th frame image:

Use the pixel value _xj,t+1 of pixel j in the t+1th frame image to match the Gaussian model i. The matching method is: if If it matches, it doesn't match. δ is a constant greater than 0, usually set δ=2.5.

S205: Determine whether the matching is successful, go to step S206, otherwise go to step S207.

S206: Update the Gaussian model i, enter step S210, the update formula is:

Among them, α is the preset model learning rate, and the value range is 0<α<1, respectively represent the weight of the i-th Gaussian model in the mixed Gaussian model of pixel j after image training in the t-th frame and the t+1-th frame, respectively represent the mean value of the i-th Gaussian model in the mixed Gaussian model of the pixel j after image training in the t-th frame and the t+1-th frame, Represent the variance of the i-th Gaussian model in the mixed Gaussian model of pixel j after image training in the t-th frame and the t+1-th frame, respectively.

S207: Determine whether i=3, if not, go to step S208, otherwise go to step S209.

S208: Let i=i+1, return to step S204.

S209: instead of updating the last Gaussian model, go to step 210:

The update formula is:

S210: Weight normalization:

After each update, the weights of the three Gaussian models To normalize, let

S211: Determine whether t=N-1, if not, go to step S212, otherwise go to step S213.

S212: Let t=t+1, return to step S202.

S213: The training is completed, and the mixed Gaussian model is obtained, and the mean value of the trained mixed Gaussian model is Variance is Weight is

In practical applications, if the surveillance video image is a grayscale image, the pixel value is a grayscale value, and only one mixed Gaussian model is required for each pixel. If it is a color image, then there are three channels of RGB, then each pixel There are three mixed Gaussian models, corresponding to the three channels of RGB.

S102: Generate a foreground binary image of the detected image:

After establishing the mixed Gaussian model of the background, it is necessary to perform foreground target detection on the detection picture frame. For the detection image in the video stream of fire exit monitoring, the foreground binary image of the frame image is obtained by using the background subtraction method according to the mixed Gaussian model.

Fig. 3 is a flow chart of foreground binary image generation. As shown in Figure 3, the foreground binary image generation includes the following steps:

S301: Select B _j Gaussian models for comparison:

The mixed Gaussian model of pixel j actually describes the probability distribution of pixel values in the time domain. In order to determine which Gaussian models in the mixed Gaussian model are generated by the background, it is necessary to divide the three Gaussian models of pixel j into Arrange in descending order, take the first B _j Gaussian models as the background distribution, the calculation formula of B _j is:

T represents a preset threshold, which means the proportion of background pixels, and the value is 0.8 in this embodiment.

S302: Match the B _j Gaussian models.

Match the pixel value x' _j of pixel point j in the detected image to B _j Gaussian models, and the matching method is the same as step S204.

S303: Determine whether any Gaussian model is successfully matched, if any one is successfully matched, go to step S304, otherwise go to step S305.

S304: Determine that the pixel point j is a background pixel, and set the value of the corresponding pixel point of the binary image to 0.

S305: Determine that pixel j is a foreground pixel, and set the value of the corresponding pixel in the binary image to 1.

In a color image, since each pixel has a Gaussian mixture model with three channels, it may appear that the determination results of foreground pixels in the three channels are different. Generally, if any one of the three channels is judged to be a foreground pixel, the pixel is considered to be a foreground pixel, and if all three channels are judged to be a background pixel, the pixel is considered to be a background pixel.

S103: Judging the authenticity of the foreground block:

Due to the change of illumination in the surveillance image, false foreground objects may appear in the foreground binary image obtained by background subtraction. In order to eliminate the influence of the transient sudden change of illumination on the detection result, the present invention introduces a histogram of oriented gradient (Histogram of Oriented Gradient, HOG) feature. The HOG feature mainly describes the edge and gradient information of the image. This feature is a feature descriptor used for object detection in computer vision and pattern recognition. It forms a gradient direction histogram by calculating and counting the gradient information of the image area. The normalization of this feature in the calculation process ensures the invariance of the optical transformation of the descriptor, and the edge and gradient information of the image are basically unchanged when the illumination changes, so the use of HOG features can well eliminate the impact of sudden illumination changes on detection. Therefore, the present invention uses the distance between the background and the HOG feature vector of the current frame image to judge the authenticity of the foreground target.

Fig. 4 is a flow chart of foreground block authenticity judgment. As shown in Figure 4, the specific method of using HOG features to judge the authenticity of the foreground block in the foreground binary image is as follows:

S401: Analyze and obtain the connected domain of the foreground binary image:

Connected domain analysis is performed on the foreground binary image obtained in step S102, the connected areas are classified into one category and the position information of the four boundaries is extracted to obtain the circumscribed rectangular coordinates of each connected domain.

Due to the presence of noise in the image, generally when performing connected domain analysis, the connected domains that are too small will be discarded directly, and only the connected domains whose size is greater than the threshold are kept, and these connected domains can be divided into blocks and the HOG feature vectors can be extracted.

S402: extract the background image:

Since the mixed Gaussian model of each pixel is composed of several Gaussian models, we need to extract a value that best represents the background to form the background. According to the Gaussian mixture model, the pixel mean value of the Gaussian model with the largest weight among the respective Gaussian models of each pixel point is used as the pixel value of the background pixel point to form a background image.

For the case where the surveillance video image is a grayscale image, then the weight of the three Gaussian models is extracted from the three Gaussian models when extracting the representative background pixel value. If the surveillance video image is a color image, then three Each of the Gaussian models extracts the largest weight, and then composes a color background image.

S403: Acquiring a grayscale image:

Acquire the detection image and the grayscale image of the background image obtained in step S402.

It can be seen that if the surveillance video image itself is a grayscale image, then the background image obtained in step S402 is naturally also a grayscale image. If it is a color surveillance video image, it is necessary to perform grayscale image conversion on the detection image and the background image.

S404: Extract two HOG feature vectors of the connected domain:

For each connected domain, the corresponding gray-scale image blocks are extracted from the gray-scale images of the detection image and the background image according to the coordinates of its circumscribed rectangle, and the HOG feature vectors K ₁ =(k ₁₁ ,k ₁₂ ,...,k _1M ) and K ₂ =(k ₂₁ ,k ₂₂ ,...,k _2M ), where M represents the number of elements in the feature vector.

In this embodiment, the specific method of extracting the HOG feature vector is as follows: firstly, divide the image block into blocks. Since the sizes of the bounding rectangles of connected domains are all the same, the present invention divides image blocks by adopting a method of fixing the number of blocks and adaptively changing the number of pixels in the blocks. The specific method of block division is: record the number of horizontal pixels of the circumscribed rectangle as X, the number of vertical pixels as Y, set the horizontal block parameter P and the vertical block parameter Q, P and Q are both natural numbers, then the horizontal pixels in each block quantity number of vertical pixels Represents rounding down, moving the block horizontally from the origin, and moving the step horizontally Reach the horizontal edge and then move vertically, the vertical movement step is This loops until the image blocks are divided into blocks. It can be seen that the number of finally obtained blocks is (2M _X -1)×(2M _Y -1). The gradient distribution in each block is counted, and finally the HOG feature vector is formed.

S405: Calculate the Euclidean distance of two HOG feature vectors:

For the two HOG feature vectors of each connected domain obtained in step S404, the Euclidean distance D of the two feature vectors is calculated. The calculation formula is:

S406: Determine whether the Euclidean distance D is greater than the preset threshold T _D , if yes, go to step S407, otherwise go to step S408.

S407: Determine that the connected domain is a real foreground block.

S408: Determine that the connected domain is a pseudo foreground block;

S104: judging abnormal conditions:

If the result analyzed in step S103 has no connected domains, or each connected domain is a false foreground block, then there is no abnormality in the frame detection image, if at least one connected domain is a real foreground block, then it is judged that the frame has abnormalities Happening. In practical applications, there may be pedestrians passing through the safety passage for a short time or the safety door is temporarily opened. Therefore, when an abnormal situation is found in a single frame image, it cannot be determined that there is a safety problem in the fire exit, and it is necessary to wait for a period of time to rule out the short-term abnormal situation. The invention is realized by continuously judging M frames of detection images, that is, if there are abnormalities in the continuous M frames, and M is the preset number of frames, then there is a safety problem in the fire exit, and a warning or alarm is required, otherwise the fire exit is safe.

During the long-term fire exit safety detection process, the scene will change due to changes in the lighting and other environments. In order to better adapt to these changes, a mixed Gaussian model of detection images to pixels can be used after each frame detection is completed. Make learning updates. Since the update of the foreground target pixels will cause the blockages that stay in the image for a long time or the normally closed fire doors that are opened to be updated into the background model so that no abnormalities can be detected, the mixed Gaussian model of the foreground target pixels cannot Update, that is, the present invention only updates the background pixels. Its update method is the same as the update method of the training sample to the mixed Gaussian model in step S101, that is, the three Gaussian models are first sorted, and then the pixel values of the background pixels are matched. Once the matching is successful, the Gaussian model is updated. If none of the three Gaussian models can be successfully matched, the last Gaussian model is replaced and updated with the pixel value of the pixel point, and the weights of the three Gaussian models are normalized after the update is completed.

In the online update process, in order to prevent the update process from making the standard deviation of the Gaussian model too small and cause the foreground detection to be too sensitive, a minimum value can be set for the standard deviation. When the calculated standard deviation is smaller than the minimum value, the standard deviation is set to is the minimum value, otherwise no operation is performed and the original value remains. In this embodiment, the minimum value is set to 5.

Obviously, when the channel monitoring image is a grayscale image, only one mixed Gaussian model is updated for each pixel. If it is a color image, the mixed Gaussian model of the three channels of RGB needs to be updated.

It can be seen that when updating online, for each pixel, it is automatically judged whether it should be updated according to its attributes. The pixels belonging to the foreground keep the original model unchanged, and only the pixels belonging to the background are updated, which is also the foreground. The target area is not updated but only the background is updated, which not only updates the scene necessary, but also avoids the error that the foreground target appears for too long and gradually blends into the background.

In practical applications, there are some very small amount of false detections in the safety detection of fire exits. For example: lighting changes cause false foreground, but this part of false foreground is not eliminated by the foreground target confirmation step and is mistakenly considered as an obstruction detected or a normally closed safety door is opened. When updating the mixed Gaussian model online, since the foreground (although it is a pseudo foreground target at this time) non-updating strategy is adopted, the background model of this part of the pseudo foreground target will not learn to update, and the current The image also basically remains unchanged. At this time, this part of the false foreground target area caused by lighting will always be wrongly detected as an abnormal state, and cannot be automatically restored to a normal state. Although this part of false detection rarely occurs, this situation will cause the detection system to continue to alarm and cannot return to normal, causing very bad effects. So this situation has to be addressed. An artificial flag can be added to the detection system to control the execution of the program and restore it to a normal state. When this kind of false detection occurs, the staff finds that the alarm is a false detection, and then manually sets the restart control flag to the restart state, deletes the original background mixed Gaussian model, and re-extracts N images from the current surveillance video stream As a training sample, the mixed Gaussian model of the background is established again, and the security detection is performed based on the new mixed Gaussian model, and the detection result will return to the normal state.

In order to verify the technical effect of the present invention, a series of experiments are carried out on real fire escape monitoring video sequences. In order to conveniently count the detection rate and false detection rate, the normal non-blocking fire exits and normally closed fire doors are normally closed as positive samples; the blocked fire exits and normally closed fire doors are opened as negative samples.

Next, 3 scenes are intercepted from the surveillance video of the fire exits captured by the same camera at 3 different times for experimental verification. Scene 1 is a normal non-blocking situation, Scene 2 is a sudden change in illumination, and Scene 3 is a blockage of fire exits. . Figure 5 is a normal detection image of scene 1. FIG. 6 is a background image obtained according to the background mixture Gaussian model at the time shown in FIG. 5 . Fig. 7 is the foreground binary image obtained in scene 1. As shown in Figures 5 to 7, since Scene 1 is a normal non-blocking situation, the obtained foreground binary image in Figure 7 has no foreground, so there is no foreground connected domain in the connected domain analysis result, so the judgment result is no abnormality Happening.

Fig. 8 is the detection image of scene 2 illumination sudden change. FIG. 9 is a background image obtained by the time background mixture model shown in FIG. 8 . Figure 10 is the foreground binary image obtained in scene 2. As shown in Figure 10, due to the sudden change in illumination, there are spots in the detection image, so there is a foreground area in the foreground binary image obtained by background subtraction.

Figure 11 is the grayscale image of the detection image and the background image extracted after the connected domain analysis of scene 2, where Figure 11(a) is the grayscale image of the detection image, and Figure 11(b) is the grayscale image of the background image. FIG. 12 is a histogram of the HOG feature vectors for scene 2. As shown in Figure 12, the HOG feature vectors of the detection image and the background image in scene 2 are very close, and the Euclidean distance is smaller than the preset threshold, so it is judged that the connected domain is a pseudo foreground, so the judgment result of scene 2 is no abnormality Happening.

Figure 13 is the detection image of channel blockage in scene 3. As shown in Figure 13, in scene 3, there are pedestrians passing through the fire escape, which blocks the passage. FIG. 14 is a background image obtained by the time background mixture model shown in FIG. 13 . Fig. 15 is the foreground binary image obtained in scene 2. As shown in Figure 13, due to pedestrians passing by, there is a foreground area in the foreground binary image obtained by background subtraction.

Fig. 16 is the grayscale image of the detection image and the background image extracted after the connected domain analysis of scene 3, wherein Fig. 16(a) is the grayscale image of the detection image, and Fig. 16(b) is the grayscale image of the background image. FIG. 17 is a histogram of the HOG feature trace of scene 3. As shown in Figure 17, in scene 3, the difference between the HOG feature vectors of the detection image and the background image in scene 2 is relatively large, and its Euclidean distance is greater than the preset threshold, so it is determined that the connected domain is the real foreground, so the scene 3 The judgment result is that there is an abnormal situation. However, since the pedestrian will pass through the fire exit in a short period of time, it will not cause long-term blockage, and the continuous frame rate of the abnormal situation is less than the preset frame rate threshold, so no warning or alarm will be generated.

Next, statistics are made on the detection results of 6 different video scenes, including sudden changes in illumination, abnormal opening of normally closed fire doors, and blockage of fire exits.

Fig. 18 is a schematic diagram of a scene with a sudden illumination change, where Fig. 18(a) is a detection image, Fig. 18(b) is a background image, and Fig. 18(c) is a detection result diagram.

Fig. 19 is a schematic diagram of a fire door opening scene 1, wherein Fig. 19(a) is a detection image, Fig. 19(b) is a background image, and Fig. 19(c) is a detection result diagram.

Fig. 20 is a schematic diagram of a fire door opening scene 2, wherein Fig. 20(a) is a detection image, Fig. 20(b) is a background image, and Fig. 20(c) is a detection result diagram.

Fig. 21 is a schematic diagram of scene 1 of fire exit blocking scene, wherein Fig. 21(a) is a detection image, Fig. 21(b) is a background image, and Fig. 21(c) is a detection result diagram.

Fig. 22 is a schematic diagram of scene 2 of fire exit blocking scene, wherein Fig. 22(a) is a detection image, Fig. 22(b) is a background image, and Fig. 22(c) is a detection result diagram.

Fig. 23 is a schematic diagram of the outdoor fire exit scene 2, wherein Fig. 23(a) is a detection image, Fig. 23(b) is a background image, and Fig. 23(c) is a detection result diagram.

Table 1 is a statistical table of quantitative detection results of 6 scene surveillance videos.

Scenes Detection rate False detection rate light mutation scene >95% <2% Fire door opening scene 1 >95% <2% Fire door opening scene 2 >95% <2% Fire exit blocking scene 1 >95% <2% Fire exit blocking scene 2 >95% <2% outdoor fire exit scene >85% <10%

Table 1

It can be seen from Table 1 that since the previous 5 video scenes are indoors, various interferences are less, relatively simple, and the detection effect is very good. It can be seen that the present invention can effectively overcome the influence of illumination changes, and is very accurate in detecting abnormal opening of fire doors and blockage of fire exits. For the sixth outdoor fire exit scene, the light changes significantly, and there are many interferences from various people and vehicles, which is relatively complicated, and the detection effect is slightly reduced, but it still maintains a good detection effect.

Although the illustrative specific embodiments of the present invention have been described above, so that those skilled in the art can understand the present invention, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those of ordinary skill in the art, As long as various changes are within the spirit and scope of the present invention defined and determined by the appended claims, these changes are obvious, and all inventions and creations using the concept of the present invention are included in the protection list.

Claims (4)

1. A fire fighting access safety detection method based on self-adaptive background learning is characterized by comprising the following steps:

s1: the method comprises the following steps of taking the first N frames of images of a fire fighting channel monitoring video stream as training samples, training to obtain a background mixed Gaussian model mixed by three Gaussian models, and specifically training the background mixed Gaussian model to obtain the background mixed Gaussian model, wherein the training steps comprise:

s1.1: initializing a Gaussian mixture model, wherein the Gaussian mixture model 1 is initialized according to the 1 st frame of image, and the expression is as follows:

wherein,representing the mean value of pixel points j after the 1 st frame of image training in the Gaussian model 1, wherein the value range of j is j =1,2, …, M is the number of pixel points in the image, and x is _j,1 Representing the pixel value of pixel point j in the 1 st frame image,representing the variance, sigma, of the pixel point j after the 1 st frame image training in the Gaussian model 1 _init Is a pre-set initial variance of the measured signal,represents the weight, omega, of the pixel point j after the 1 st frame image training in the Gaussian model 1 _init Is a preset initial weight;

the gaussian models 2,3 are initialized with the expression:

respectively representing the average value of pixel points j after the 1 st frame of image training in the Gaussian models 2 and 3, represents the variance of the pixel point j after the 1 st frame image in the Gaussian models 2 and 3 is trained,representing the weight of a pixel point j after the 1 st frame of image in the Gaussian models 2 and 3 is trained;

s1.2: the method for training the frame t +1 image comprises the following steps of sequentially training a mixed Gaussian model by using the frame t 2,3 and the frame t … N image, and obtaining the mixed Gaussian model, wherein the method for training the frame t +1 image comprises the following steps:

three Gaussian models of pixel point j, t =1,2, … and N-1, obtained through the previous t frame image are processed according to the methodI =1,2,3, using the pixel value x of the pixel point j in the t +1 frame image _j,t+1 And sequentially matching each Gaussian model, wherein the matching method comprises the following steps: if it is notWherein δ is a constant greater than 0, then it is considered to be matched, otherwise it is not matched;

once a Gaussian model is successfully matched, the Gaussian model takes the pixel value x _j,t+1 Updating, keeping other Gaussian models unchanged, and updating the formula as follows:

wherein alpha is a preset model learning rate, the value range is more than 0 and less than 1,respectively representing the weight of the ith Gaussian model in the mixed Gaussian models of the pixel point j after the image training of the t frame and the t +1 frame,respectively representing the mean value of the ith Gaussian model in the mixed Gaussian models of the pixel point j after the t frame and the t +1 frame are trained,respectively representing the variance of the ith Gaussian model in the mixed Gaussian models of the pixel point j after the image training of the t frame and the t +1 frame;

if the pixel value x _j,t+1 If the three Gaussian models are not matched, the last Gaussian model is used according to the pixel value x _j,t+1 Performing replacement updating, wherein an updating formula is as follows:

after the updating is completed, the weights of the three Gaussian models are usedPerforming normalization toThe mean value of the Gaussian mixture model after training is recorded asVariance ofThe weight is

S2: for each frame of detection image in the fire fighting channel monitoring video stream, obtaining a foreground binary image of the frame of image by adopting a background subtraction method according to a Gaussian mixture model, and the method specifically comprises the following steps:

s2.1: pressing the three Gaussian models of the pixel point jPerforming descending order, taking the first B _j Distribution of a Gaussian model as background, B _j The calculation formula of (c) is:

t represents a preset threshold value;

s2.2: detecting the pixel value x 'of the pixel point j in the image' _j To B _j Matching the Gaussian models, if any Gaussian model is successfully matched, judging that the pixel point j is a background pixel, setting the value of the pixel point corresponding to the binary image to be 0, otherwise, judging that the pixel point j is a foreground pixel, and setting the value of the pixel point corresponding to the binary image to be 1;

s3: the authenticity of a foreground block in a foreground binary image is judged by using the HOG characteristics, and the method specifically comprises the following steps:

s3.1: carrying out connected domain analysis on the foreground binary image obtained in the step S2, classifying the areas connected into blocks into one class, extracting position information of four boundaries, and obtaining circumscribed rectangular coordinates of each connected area;

s3.2: taking the pixel mean value of the Gaussian model with the maximum weight in the three Gaussian models of each pixel point as the pixel value of the background pixel point to form a background image;

s3.3: for each connected region, extracting corresponding gray image blocks from the gray images of the detection image and the background image respectively according to external rectangular coordinates of the connected region, extracting HOG feature vectors of the two gray image blocks respectively, calculating Euclidean distance D of the two feature vectors, and if D is larger than a preset threshold value T _D If yes, judging the connected region as a real foreground block, otherwise, judging the connected region as a false foreground block;

s4: if each connected region in the step S3 is a pseudo foreground block, the frame detection image has no abnormal condition, and if at least one connected region is a real foreground block, the frame is judged to have the abnormal condition; if the continuous M frames have abnormal conditions, and M is a preset frame number, the fire fighting channel has a safety problem, otherwise, the fire fighting channel is safe;

s5: after the detection is finished, the Gaussian mixture model is learned and updated according to the detection image, and the specific method comprises the following steps: first, the three Gaussian models are calculated according toThe values are arranged in a descending order, then the values are matched with the pixel values of the background pixel points in sequence, the Gaussian model is updated once the matching is successful, if the matching of the three Gaussian models is not successful, the last Gaussian model is replaced and updated by the pixel value of the pixel point, and the weights of the three Gaussian models are normalized after the updating is completed.

2. A fire fighting access safety detection method based on adaptive background learning according to claim 1, wherein in step S3.3, the extraction method of the HOG feature vector is as follows:

recording the number of transverse pixels of the circumscribed rectangle as X and the number of longitudinal pixels as Y, setting a transverse blocking parameter P and a longitudinal blocking parameter Q, wherein both P and Q are natural numbers, and then the number of transverse pixels in each blockNumber of vertical pixels Represents rounding down, moving the block laterally from the origin, moving the step size laterallyReach the transverse edge and move longitudinally by the step length ofThe steps are circulated until the image blocks are partitioned; make statistics of eachAnd distributing the gradients in the block to finally form the HOG characteristic vector.

3. The fire fighting access safety detection method based on adaptive background learning of claim 1, wherein in the learning update process of step S5, a minimum value is set for the standard deviation, and when the calculated standard deviation is smaller than the minimum value, the standard deviation is set to the minimum value.

4. The fire fighting access safety detection method based on adaptive background learning according to claim 1, further comprising processing false detection, and the specific method is as follows: and when false detection occurs, deleting the original background Gaussian mixture model, returning to the step S1 again to extract N images from the current monitoring video stream as training samples, establishing the background Gaussian mixture model again, and performing safety detection based on the new Gaussian mixture model.