Smoke and fire detection method and system with memory module

Abstract

The invention relates to a smoke and fire detection method and system with a memory module, and aims to solve the problem of high false alarm rate caused by interference sources such as factory smoke and cooking smoke. The method comprises the steps of performing smoke and fire detection on a to-be-detected image and generating an initial detection frame, wherein the initial detection frame comprises a target position, a category and confidence; calculating a homography matrix between the to-be-detected image and a standard sample, and migrating a preset false detection prior frame to a to-be-detected image coordinate system through the homography matrix; according to a preset coverage ratio threshold value, by comparing the space intersection area of the initial detection frame and the migrated false detection prior frame, fixed smoke and fire or common smoke and fire are judged and marked; and when at least one initial detection frame marked as common smoke and fire is detected, an alarm signal is triggered. The method effectively reduces the false alarm rate, improves the detection reliability, and is suitable for the field of public safety and disaster early warning.

Description

Technical Field

[0001] The present invention relates to the fields of computer vision and image processing, and in particular to a smoke and fire detection method and system with a memory module for fire early warning. Background Art

[0002] In the fields of public safety and disaster prevention, fire and smoke detection technology has long been a hot topic of research and application. With the acceleration of urbanization and the growing need to protect forest resources, accurate and timely detection of fire and smoke, while avoiding false alarms, is crucial for safeguarding life and property and maintaining ecological stability. Existing fire and smoke detection technologies primarily rely on image recognition and sensor monitoring, analyzing image features such as color, shape, and texture, or monitoring parameters such as temperature and gas composition to determine the presence of fire and smoke.

[0003] However, in practical applications, smoke and fire detection faces a thorny challenge: factory smoke, residential smoke, and nighttime lights, among other things, can often share similar or identical characteristics with real fireworks after imaging. For example, factory and cooking smoke can resemble early-stage fireworks in color and dynamic changes, while lights can be misidentified as flames under certain lighting conditions. Traditional smoke and fire detection methods often struggle to effectively distinguish these similar characteristics, resulting in a high incidence of false alarms.

[0004] Frequent false alarms not only waste significant manpower, material resources, and time to verify false alarms, but can also reduce the reliability of detection systems, potentially leading to the risk of missed alarms when actual fireworks incidents occur. Therefore, reducing the false alarm rate in fireworks detection has become a key technical issue that urgently needs to be addressed.

[0005] Currently, several improved detection algorithms have been proposed to address these issues, such as increasing feature dimensions and optimizing classifiers to improve detection accuracy. However, these methods mostly focus on analyzing real-time detection data while neglecting the use of historical detection data. In reality, in specific detection areas, such as those surrounding factories and residential areas, interference sources such as factory smoke, cooking smoke, and lights appear with a certain regularity and repetitiveness. If we can analyze and summarize the false detections in historical detection results and establish corresponding false detection models, we can more effectively filter out these interference factors in subsequent detections.

[0006] Based on the above situation, the present invention aims to provide a fireworks detection method and system with a memory module. By incorporating the memory module, cluster analysis is performed on historical detection results to generate a priori false positives, which are then applied to subsequent detection processes. This method fully utilizes the information contained in historical data and can effectively identify interference sources with similar characteristics to fireworks, thereby reducing false alarm rates and improving the reliability and practicality of the fireworks detection system. This has important practical significance for improving public safety and reducing resource waste. Summary of the Invention

[0007] The purpose of the present invention is to solve the technical problem of high false alarm rate caused by the fact that the characteristics of interference sources such as factory smoke, residential cooking smoke and night lights are similar to real fireworks, and the existing detection methods do not effectively use historical data to identify repeated false detection areas.

[0008] In order to achieve the above-mentioned purpose, the present invention adopts the following technical means:

[0009] The present invention provides a method for detecting fireworks with a memory module, comprising the following steps:

[0010] S1. Perform fireworks detection on the image to be detected and generate an initial detection frame, which includes the target location, category and confidence level;

[0011] S2. Calculate the homography matrix between the image to be detected and the standard sample, and migrate the preset false detection prior frame to the coordinate system of the image to be detected through the homography matrix;

[0012] S3. According to a preset coverage ratio threshold, by comparing the spatial intersection area of ​​the initial detection frame and the migrated false detection prior frame, it is determined and marked as fixed fireworks or ordinary fireworks;

[0013] S4. When at least one initial detection frame marked as ordinary fireworks is detected, an alarm signal is triggered.

[0014] In the above scheme, step 1 includes the following steps:

[0015] Performing preprocessing on the image to be detected, wherein the preprocessing includes image enhancement and denoising operations;

[0016] Use a deep learning-based fireworks detection algorithm to detect the preprocessed image and generate an initial detection frame that includes the target location, target category, and detection confidence.

[0017] The initial detection frames are screened for confidence, and detection frames whose detection confidence is lower than a preset minimum threshold are removed, where the preset minimum threshold is 0.3.

[0018] In the above scheme, step 2 includes the following steps:

[0019] Using a feature matching algorithm to extract and match feature point pairs between the image to be detected and the standard sample image to obtain a sufficient number of corresponding feature points;

[0020] The RANSAC algorithm is used to calculate the homography matrix H between the image to be detected and the standard sample. The homography matrix is ​​a 3×3 transformation matrix, and its specific form is as follows:

[0021]

[0022] The coordinates of the preset false detection prior frame in the standard sample image are transferred by the homography matrix H according to the formula:

[0023]

[0024] Calculate its transformed coordinates (x′, y′) in the image to be detected, where:

[0025] H is the homography matrix;

[0026] (x, y) is the coordinate of the false detection prior box in the standard sample image;

[0027] (x′, y′) is the coordinate in the image to be detected after transformation;

[0028] After performing the above homography transformation on the four vertices of the false detection prior frame, if the transformed coordinates exceed the range of the image to be detected, the excess part is cropped or marked.

[0029] In the above scheme, step 3 includes the following steps:

[0030] For each initial detection frame, calculate the intersection area between it and all false detection prior frames. lntersection And the area of ​​the initial detection frame itself Test , and calculate the coverage score Cover ratio :

[0031]

[0032] For each initial detection frame, select the maximum coverage ratio score between it and all false detection prior frames;

[0033] If the maximum coverage ratio score is greater than or equal to a preset threshold, the initial detection frame is marked as fixed fireworks; if it is less than the preset threshold, it is marked as ordinary fireworks.

[0034] In the above scheme, step 4 includes the following steps:

[0035] After all initial detection frames are marked, the detection frames marked as ordinary fireworks are filtered out;

[0036] If the screening results contain at least one common fireworks detection frame, the corresponding image to be detected, the location information of the detection frame and the detection time will be reported to the monitoring center; if the screening results do not contain any common fireworks detection frame, no alarm or reporting process will be triggered.

[0037] In the above scheme, the calculation process of the false detection prior box in step 2 is as follows:

[0038] S201. Obtaining historical detection images and their corresponding detection frame information, wherein the detection frame information includes target coordinates, detection time, and detection confidence;

[0039] S202. Select a historical detection image as a standard sample image, and calculate the homography matrix H between the remaining historical detection images and the standard sample image;

[0040] Furthermore, the matching point pairs between the standard sample and the historical detection image are calculated through image matching algorithms (such as SIFT, SURF, Efficient LoFTR, etc.). Then, the RANSAC algorithm is used to randomly collect 8 pairs of points from the matching point pairs to calculate the transformation matrix. The transformation matrix with the smallest error after the transformation of other point pairs is the homography matrix H.

[0041] S203. Migrating the detection frames in the remaining historical detection images to the standard sample image coordinate system through the homography matrix H;

[0042] S204. Perform density clustering on all the migrated detection frames to generate clusters, and calculate the prior frame parameters corresponding to each cluster based on the following formula:

[0043]

[0044]

[0045]

[0046]

[0047]

[0048]

[0049] W i =e -λd

[0050] Where W i is the time decay weight for each box, x i and y i is the center coordinate of each box, N is the number of detection boxes in the cluster, Xpre and Y pre is the center coordinate of the prior frame, σ x and σ y is the standard deviation of the center coordinates in the x and y directions, Width pre and Height pre is the width and height of the prior frame, d is the number of days from the image time to the current date, λ is the attenuation factor, width i , height i is the width and height of the i-th detection box.

[0051] S205. Calculate the confidence of each prior box in:

[0052] N T is the total number of historical detection frames, k and a are preset adjustment parameters;

[0053] If Conf is greater than the preset threshold, the priori box is determined to be a false detection priori box.

[0054] The above solution also includes the steps of storing new history frames:

[0055] S301. Extract two historical detection images and their corresponding detection frame information at different time periods every day and store them in the historical detection record. The detection frame information includes the detection timestamp, target coordinates and detection confidence.

[0056] S302. At the beginning of each day's detection task, steps S201 to S205 are triggered to generate a false positive priori frame for that day. The clustering calculation of the false positive priori frame uses a time decay weight w. i The historical detection frames are weighted and the false detection prior frames are cleared after the detection task of the day is completed.

[0057] The present invention also provides a smoke and fire detection system with a memory module. The system is stored in a storage medium. When a processor executes the system, the smoke and fire detection method with a memory module is implemented.

[0058] Because the present invention adopts the above technical means, it has the following beneficial effects:

[0059] 1. This invention addresses the technical problem of repeated false alarms caused by fixed interference sources in traditional fire and smoke detection by introducing density cluster analysis of historical detection frames and the migration of false positive prior frames (corresponding to steps 2.4 and 1.2). It effectively filters out regular interference such as factory smoke and residential cooking smoke, significantly reducing the system's false alarm rate.

[0060] 2. This invention solves the technical problem of filtering failure caused by insufficient timeliness of historical data by calculating time-decayed weights and dynamically updating the prior frame (corresponding to steps 2.4 and 3.1), thereby achieving the effect of adapting to dynamic environmental changes and improving the generalization ability of the false detection model.

[0061] 3. This invention solves the technical problem of traditional methods being unable to distinguish between new fires and historical false positives by calculating the coverage ratio between the detection frame and the false positive prior frame (corresponding to step 1.3). This effectively reduces the manual verification cost of repetitive false positives while ensuring the detection rate of real fires.

[0062] 4. The present invention solves the limitation problem of existing technology relying on manually labeled historical data through the technical means of unsupervised clustering and automatic weight updating (corresponding to steps 2.4 and 3.2), realizes system self-optimization and iteration, significantly improves deployment efficiency and reduces operation and maintenance complexity. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 It is a simplified flow chart of the present invention;

[0064] Figure 2 This is the calculation process of the false detection prior frame of the present invention. DETAILED DESCRIPTION

[0065] The following is a detailed description of the embodiments of the present invention. Although the present invention will be described and illustrated in conjunction with certain specific embodiments, it should be noted that the present invention is not limited to these embodiments. On the contrary, modifications or equivalent substitutions of the present invention are intended to fall within the scope of the claims of the present invention.

[0066] In addition, in order to better illustrate the present invention, numerous specific details are given in the following detailed description. It will be understood by those skilled in the art that the present invention can also be implemented without these specific details.

[0067] The present invention aims to provide a method and system for detecting fireworks with a memory module. By clustering historical detection results, a priori false positive boxes are obtained and aligned with standard samples. A fire detection frame is then generated for the image to be detected. The false positive boxes in the standard samples are then transferred to the image to be detected through image alignment. Finally, the positional relationship between the detection frame and the priori false positive boxes is calculated to filter out fireworks that appear in the same area, reducing false positives.

[0068] Example 1

[0069] The implementation of the present invention consists of the following steps, and the detection flow chart is as follows: Figure 1 As shown:

[0070] Step 1: Perform fireworks detection on the image to be detected and filter it with the false detection prior frame

[0071] Step 1.1: Perform fireworks detection on the image to be detected and obtain the detection frame

[0072] The image to be inspected is preprocessed, including image enhancement and denoising, to improve the accuracy of subsequent detection. A fireworks detection algorithm (such as the deep learning-based YOLO and Faster R-CNN) is then used to detect the preprocessed image. These algorithms, trained on a large dataset of fireworks images, are able to accurately identify fireworks targets within the image. During the detection process, the algorithm generates a series of detection boxes, each containing information such as the target's location and category, as well as the detection confidence level. The detection results are then preliminarily screened, removing boxes with a detection confidence level below a preset minimum threshold (e.g., 0.3) and retaining detection results with higher confidence levels.

[0073] Step 1.2: Calculate the homography matrix between the image to be detected and the standard sample, and transfer the false detection prior box to the image to be detected;

[0074] Since the camera runs for a long time, factors such as wind may cause camera shaking at different times. Therefore, images taken at different times need to be aligned.

[0075] First, a feature matching algorithm (such as SIFT, SURF, or Efficient LoFTR) is used to extract and match feature points between the image to be tested and the standard sample image to find a sufficient number of corresponding feature point pairs. Then, the RANSAC algorithm is used to calculate the homography matrix H between the image to be tested and the standard sample image. The homography matrix is ​​a 3x3 matrix with the following form:

[0076]

[0077] It describes the projective transformation relationship from one plane to another, and can map points in one image to corresponding points in another image. After obtaining the homography matrix, the false positive prior box is transformed, migrating it from the coordinate space of the standard sample image to the coordinate space of the image to be detected. The specific formula for the homography transformation is:

[0078]

[0079] Where (x, y) is the coordinate of the false positive prior box in the standard sample image, and (x', y') is the coordinate of the transformed false positive prior box in the image to be detected. By transforming the four vertices of the false positive prior box, the position and size of the transformed false positive prior box in the image to be detected are obtained. It is important to note that during the transformation process, the image boundary restrictions must be considered. For the part of the detection box that exceeds the range of the image to be detected after the transformation, appropriate cropping or marking is performed to ensure that the transferred false positive prior box is reasonably represented in the image to be detected.

[0080] Step 1.3: Calculate the coverage ratio score between each detection frame and the false detection prior frame. If the coverage ratio is greater than the threshold, the detection frame is marked as a fixed firework, otherwise it is an ordinary firework.

[0081] For each detection box obtained in step 1.1, calculate the coverage ratio between it and each false positive prior box transferred to the image to be detected in step 1.2. The coverage ratio is calculated by dividing the intersection area of ​​the two boxes by the area of ​​the detection box, that is:

[0082]

[0083] Among them, Area (Intersection) is the intersection area of ​​the detection frame and the false detection prior frame, and Area (Test) is the area of ​​the detection frame. For each detection frame, find the maximum coverage ratio score with all false detection prior frames. Compare the maximum coverage ratio score with the preset threshold (such as 0.5). If it is greater than or equal to the threshold, it means that the area corresponding to the detection frame has a large overlap with the historical false detection prior frame area, and it is likely to be a fixed false detection source (such as long-standing heating equipment, fixed reflective areas, etc.). Therefore, the detection frame is marked as fixed fireworks; conversely, if the maximum coverage ratio score is less than the threshold, it means that the area corresponding to the detection frame is a newly appeared fireworks target that is unrelated to the historical false detection area, and it is marked as ordinary fireworks.

[0084] Step 1.4: If there are still ordinary fireworks remaining in the image to be detected, report it; otherwise, do not report it

[0085] After all detection frames are marked in step 1.3, select the detection frames marked as ordinary fireworks. If at least one ordinary fireworks detection frame exists, it indicates that there may be real fireworks targets in the image to be detected, not fixed false detection sources. Relevant information (such as the detected image, detection frame location, detection time, etc.) needs to be reported to the monitoring center or relevant personnel for further confirmation and processing. If no ordinary fireworks detection frames exist, that is, all detection frames are marked as fixed fireworks, it means that these detection results are likely caused by historical false detection sources and do not need to be reported, thereby reducing the false alarm rate.

[0086] Step 2: Update the false positive prior frame. The calculation process of the false positive prior frame is as follows: Figure 2 shown

[0087] Step 2.1: Get the historical detection image and its corresponding detection frame

[0088] In the historical data management module of the fireworks detection system, historical detection images are batch retrieved from a long-term storage database through a data interface. These images cover monitoring data from the same monitoring point at different time points (including various environmental conditions such as daytime, nighttime, sunny, and rainy days). The corresponding detection frame information includes detailed parameters such as the specific coordinates of the target in the image (top left and bottom right corners), the target's width and height, the detection timestamp, and the detection confidence level.

[0089] Step 2.2: Select a historical detection image as a standard sample and match it with the remaining images to obtain the homography matrix.

[0090] From the verified historical detection images, manually or automatically select a representative image as the standard sample based on factors such as image clarity, scene representativeness, and detection frame integrity. Calculate the homography matrix between the remaining samples and the standard sample using the same method as in step 1.2.

[0091] Step 2.3: Perform homography transformation on the detection boxes in the remaining images and migrate them all to the standard sample:

[0092] For each historical detection image of a non-standard sample, obtain the corresponding detection frame coordinate information. Based on the homography matrix obtained in step 2.2, perform a homography transformation on the four vertex coordinates (top left and bottom right corner coordinates) of each detection frame. Migrate all historical detection frames to the standard sample.

[0093] Step 2.4: Perform density clustering on all detection frames in the standard sample to obtain the prior frame:

[0094] Extract all the detection box information (including center x and y coordinates, width, height, and detection time) from the standard sample in step 2.3 to form a data set. Use a density clustering algorithm (such as the DBSCAN algorithm) to perform cluster analysis on the center coordinates in this data set to obtain clusters.

[0095] Each cluster corresponds to a priori box. The position and size of the priori box are determined by statistical analysis of the detection boxes within the cluster. The weighted center coordinates, average width, and height of the detection boxes in the cluster are taken. Through density clustering, similar detection boxes can be grouped together to form representative priori boxes. These priori boxes reflect the distribution and characteristics of false positives that may have occurred in historical detections. The calculation formula for the priori box of each cluster is as follows:

[0096]

[0097]

[0098]

[0099]

[0100]

[0101]

[0102] W i =e -λd

[0103] Where W i is the time decay weight for each box, x i and y i is the center coordinate of each box, N is the number of detection boxes in the cluster, X pre and Y pre is the center coordinate of the prior frame, σ x and σ y is the standard deviation of the center coordinates in the x and y directions, Width pre and Height pre is the width and height of the prior frame, d is the number of days from the image time to the current date, λ is the attenuation factor, width i , height i is the width and height of the i-th detection box.

[0104] Step 2.5: Calculate the confidence of each prior frame. If it is greater than the threshold, it is a false positive prior frame.

[0105] For each prior box obtained in step 2.4, calculate its confidence. The confidence calculation is based on factors such as the number of detection boxes in the cluster where the prior box is located, the distribution of detection time, and the consistency of the detection algorithm. Specifically, the confidence can be calculated using the following formula:

[0106]

[0107] Where N is the number of detections of the detection box in the cluster ( Figure 2 The number of detections of the cluster box in the upper left corner is 5), N T is the number of detections of the historical detection box ( Figure 2 The total number of detections in the example is 8). After calculating the confidence of each prior frame, it is compared with the preset threshold ( Figure 2 If the confidence level is greater than the threshold, it means that the area represented by the prior box has had similar detection results many times in historical detections and has a high degree of confidence. It may be an area prone to false detections, so it is determined as a false detection prior box. Otherwise, the prior box is excluded.

[0108] Step 3: Add new history frame to the database

[0109] Step 3.1: Every day, two images and their detection results are extracted at different time periods and put into the historical detection record. Images that are farther away from the current date are assigned lower clustering weights, so that the false detection prior box is more dependent on the most recent detection images.

[0110] To ensure the timeliness and diversity of historical detection data, the system automatically extracts two representative images and their corresponding detection results from real-time detection data at different time periods each day (such as 2:00 AM, 10:00 AM, 4:00 PM, and 8:00 PM, which can be adjusted based on actual monitoring needs). During this extraction process, priority is given to images with significant environmental variations and unique detection results, thereby enriching the diversity of historical data.

[0111] Step 3.2: Trigger step 2 once every day at the beginning of the detection task to obtain the false detection prior frame of each point, and use the prior frame to filter the images detected that day.

[0112] Every day when the detection task starts, the system will automatically trigger the operation of step 2, and generate a false positive prior frame for each monitoring point (if there are multiple monitoring points) based on the latest historical detection records (including newly stored images and previous historical data). Since different weights are assigned to historical images in step 3.1, the weight factor will be taken into account when calculating the false positive prior frame in step 2.4, so that the clustering results are more inclined to recent data. The generated false positive prior frame will be stored in the system's temporary cache for use in the detection task of the day. After fireworks detection is performed on each image to be detected on that day, the false positive prior frame generated on that day is used to filter the detection results of the image to be detected according to the process of step 1, thereby realizing real-time false positive filtering of the detection data of the day and improving the accuracy and reliability of the detection results. After the detection task is completed every day, the false positive prior frame generated on that day will be cleared so that a new false positive prior frame can be regenerated the next day to ensure that each detection is based on the latest historical data.

[0113] Example 2

[0114] The present invention also provides a smoke and fire detection system with a memory module. The system is stored in a storage medium. When a processor executes the system, the smoke and fire detection method with a memory module is implemented.

Claims

1. A method for detecting fireworks with a memory module, characterized in that: The following steps are involved: S1. Perform fireworks detection on the image to be detected and generate an initial detection frame, which includes the target location, category and confidence level; S2. Calculate the homography matrix between the image to be detected and the standard sample, and migrate the preset false detection prior frame to the coordinate system of the image to be detected through the homography matrix; S3. According to a preset coverage ratio threshold, by comparing the spatial intersection area of ​​the initial detection frame and the migrated false detection prior frame, it is determined and marked as fixed fireworks or ordinary fireworks; S4. When at least one initial detection frame marked as ordinary fireworks is detected, an alarm signal is triggered.

2. The method according to claim 1, characterized in that Step 1 includes the following steps: Performing preprocessing on the image to be detected, wherein the preprocessing includes image enhancement and denoising operations; Use a deep learning-based fireworks detection algorithm to detect the preprocessed image and generate an initial detection frame that includes the target location, target category, and detection confidence. The initial detection frames are screened for confidence, and detection frames whose detection confidence is lower than a preset minimum threshold are removed, where the preset minimum threshold is 0.

3.

3. The method according to claim 1, characterized in that Step 2 includes the following steps: Using a feature matching algorithm to extract and match feature point pairs between the image to be detected and the standard sample image to obtain a sufficient number of corresponding feature points; Calculate the homography matrix H between the image to be detected and the standard sample using the RANSAC algorithm, where the homography matrix is ​​a 3×3 transformation matrix; The coordinates of the preset false detection prior frame in the standard sample image are transferred by the homography matrix H according to the formula: Calculate its transformed coordinates (x′, y′) in the image to be detected, where: H is the homography matrix; (x, y) is the coordinate of the false detection prior box in the standard sample image; (x′, y′) is the coordinate in the image to be detected after transformation; After performing the above homography transformation on the four vertices of the false detection prior frame, if the transformed coordinates exceed the range of the image to be detected, the excess part is cropped or marked.

4. The method according to claim 1, wherein Step 3 includes the following steps: For each initial detection frame, calculate the intersection area between it and all false detection prior frames. Intersection And the area of ​​the initial detection frame itself Test , and calculate the coverage score Cover ratio : For each initial detection frame, select the maximum coverage ratio score between it and all false detection prior frames; If the maximum coverage ratio score is greater than or equal to a preset threshold, the initial detection frame is marked as fixed fireworks; if it is less than the preset threshold, it is marked as ordinary fireworks.

5. The method according to claim 1, wherein Step 4 includes the following steps: After all initial detection frames are marked, the detection frames marked as ordinary fireworks are filtered out; If the screening results contain at least one common fireworks detection frame, the corresponding image to be detected, the location information of the detection frame and the detection time will be reported to the monitoring center; if the screening results do not contain any common fireworks detection frame, no alarm or reporting process will be triggered.

6. The method according to claim 1, characterized in that The calculation process of the false positive prior frame in step 2 is as follows: S201. Obtaining historical detection images and their corresponding detection frame information, wherein the detection frame information includes target coordinates, detection time, and detection confidence; S202. Select a historical detection image as a standard sample image, and calculate the homography matrix H between the remaining historical detection images and the standard sample image; S203. Migrating the detection frames in the remaining historical detection images to the standard sample image coordinate system through the homography matrix H; S204. Perform density clustering on all the migrated detection boxes to generate clusters, and calculate the prior box parameters corresponding to each cluster based on the following formula: W i =e -λd Where W i is the time decay weight for each box, x i and y i is the center coordinate of each box, N is the number of detection boxes in the cluster, X pre and Y pre is the center coordinate of the prior frame, σ x and σ y is the standard deviation of the center coordinates in the x and y directions, W pre and H pre is the width and height of the prior frame, d is the number of days from the image time to the current date, λ is the attenuation factor, width i , height i is the width and height of the i-th detection box. S205. Calculate the confidence of each prior box in: N T is the total number of historical detection frames, k and a are preset adjustment parameters; If Conf is greater than the preset threshold, the priori box is determined to be a false detection priori box.

7. The method according to claim 6, characterized in that It also includes the steps for adding a new history frame to the database: S301. Extract two historical detection images and their corresponding detection frame information at different time periods every day and store them in the historical detection record. The detection frame information includes the detection timestamp, target coordinates and detection confidence. S302. At the beginning of each day, the detection task is triggered at step S201 to S205 to generate a false positive prior frame for that day. The cluster calculation of the false positive prior frame uses a time decay weight w. i The historical detection frames are weighted and the false detection prior frames are cleared after the detection task of the day is completed.

8. A fire and smoke detection system with a memory module, characterized in that: The system is stored in a storage medium, and when a processor executes the system, the method according to any one of claims 1 to 7 is implemented.

Images