Lightweight fire detection and emergency evacuation device and method and storage medium
By combining lightweight AI models with edge computing terminals, the system enables very early and accurate detection of fires in small and medium-sized venues and dynamic evacuation route planning. This addresses the low-cost and high-precision requirements for fire detection and emergency evacuation in small and medium-sized venues, and improves emergency response efficiency and the flexibility of fire identification.
Patent Information
- Application Number
- CN202610219376.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-02-24
- Publication Date
- 2026-05-12
AI Technical Summary
Existing technologies cannot meet the needs of low-cost, high-precision fire detection and emergency evacuation in small and medium-sized venues. Traditional equipment is susceptible to environmental interference, has high model deployment costs, low model accuracy, and fixed evacuation guidance methods that cannot be dynamically adapted to fire scenarios. Detection and guidance are disconnected, resulting in low emergency response efficiency.
A lightweight AI model is adopted, which combines visual and temperature features to identify fires at the earliest stage. The model is processed in real time through an edge computing terminal to dynamically plan evacuation routes. Multi-dimensional devices are used for alarms and guidance to form a closed-loop response. The model has 1.8M parameters, training accuracy ≥99.2%, and device power consumption ≤15W.
It enables very early and accurate detection of fires in small and medium-sized venues, dynamic evacuation route planning, multi-dimensional alarms and guidance, improves emergency response efficiency, reduces deployment costs, adapts to different lighting and obstruction scenarios, and enhances the flexibility and accuracy of fire identification.
Smart Images
Figure CN122024447A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of fire prevention and emergency evacuation technology, specifically to a lightweight fire detection and emergency evacuation device, method, and storage medium. Background Technology
[0002] Small and medium-sized venues face unique challenges in fire prevention and control due to their flexible spatial layout, frequent personnel flow, and limited investment in fire protection facilities. On the one hand, traditional fire detection equipment, such as point-type smoke detectors, is easily affected by environmental interference such as oil fumes, dust, and steam, resulting in a false alarm rate as high as 10% to 15%, which leads to desensitization to alarms among venue personnel. On the other hand, existing AI image fire detection systems are mainly designed for large venues, with complex models, parameters ≥100M, high hardware requirements, the need to deploy dedicated GPU servers, and high deployment costs, with an investment of ≥50,000 yuan per venue, making it difficult to meet the low-cost and easy-to-maintain needs of small and medium-sized venues.
[0003] Emergency evacuation in small and medium-sized venues often relies on fixed indicator lights and paper plans, which have obvious limitations. Fixed indicator lights cannot dynamically adjust evacuation routes according to the location of the fire and the direction of smoke spread, and may easily lead people into dangerous areas. Paper plans lack real-time capability and cannot be quickly transmitted to people on site when a fire occurs. In addition, most small and medium-sized venues are not equipped with professional emergency broadcasting systems, and alarm information is not transmitted in a timely and accurate manner, leading to chaotic evacuation and even safety accidents.
[0004] Chinese Patent No. 202411514696X discloses an AI-based image-based fire identification and alarm method and a fire detector. This technical solution utilizes a centrally managed model library platform. The platform provider is responsible for collecting, establishing, and maintaining a fire identification model library containing various scene adaptability models. These models have undergone rigorous testing and optimization, possessing high identification accuracy and generalization ability. When a user submits a fire monitoring request, the platform can quickly analyze the user's needs, match, and recommend the most suitable fire identification model. However, this technical solution cannot flexibly adapt to the diverse needs of small and medium-sized venues, and updating the model library requires a considerable amount of time, thus its flexibility needs improvement.
[0005] The application of existing lightweight AI image recognition technology in the field of fire detection still has shortcomings. In pursuit of model simplification, many technologies use single features (such as flame color and smoke outline) for recognition, ignoring the composite features of "weak smoke + abnormal temperature" in the early stages of a fire, resulting in an accuracy rate of ≤85% for very early detection. The models lack scene adaptability and their performance degrades significantly under different lighting conditions (such as low light at night and direct strong light) and obstruction conditions (such as shelves blocking smoke). Furthermore, they are not deeply integrated with the evacuation guidance process, resulting in a disconnect between detection and guidance.
[0006] In summary, existing technologies cannot meet the needs of small and medium-sized venues, mainly due to the following shortcomings: ① High deployment cost and poor compatibility of AI detection models; ② Fire identification relies on single features, resulting in low accuracy and weak anti-interference capabilities in the very early stages; ③ Fixed evacuation guidance methods cannot dynamically adapt to fire scenarios; ④ Fragmented processes from detection to alarm to guidance, leading to low emergency response efficiency. Therefore, there is an urgent need to propose a lightweight, high-precision, low-cost, and integrated method that combines detection and guidance. Summary of the Invention
[0007] The purpose of this invention is to overcome the shortcomings of the prior art and provide a lightweight fire detection and emergency evacuation device, method and storage medium.
[0008] In a first aspect, the present invention provides a fire detection method for small and medium-sized venues, comprising the following steps: Step 101: Scene modeling and equipment deployment in small and medium-sized venues. Complete the digital modeling of the scene and the deployment of lightweight equipment to generate a scene topology map; Step 102: Training and deployment of lightweight AI fire detection model, building a lightweight AI model suitable for edge operation, and achieving accurate identification of multiple features of fires in the very early stage; Step 103: Real-time fire detection and risk assessment. The edge computing terminal processes camera image data in real time to complete fire identification and risk level determination. Step 104: Dynamic evacuation route planning and updating. Based on the scene topology map and real-time fire data, an improved path planning algorithm is used to generate the optimal evacuation route. Step 105: Multi-dimensional emergency alarm and evacuation guidance. Relying on the deployed lightweight equipment, alarms and guidance are achieved through multiple dimensions of vision, hearing and mobile terminals to form a closed-loop response. Step 106: Fire data analysis and detection model update. After evacuation, fire data is collected, analyzed, and transmitted to a lightweight AI fire detection model for training and updating to improve the accuracy of fire detection.
[0009] Furthermore, the lightweight AI fire detection model uses MobileNetV3 as its backbone network, with 1.8M model parameters.
[0010] Furthermore, the lightweight AI fire detection model further includes a feature fusion module, which extracts the RGB channel difference of flame color, LBP features of smoke texture, and regional temperature mean and mutation rate for feature fusion.
[0011] Furthermore, the training accuracy of the lightweight AI fire detection model is ≥99.2%, and the model size is further compressed to 0.8MB through INT8 quantization compression.
[0012] Furthermore, the path planning uses the formula P = D × (1 + Δ) to calculate the path cost, where P is the path cost, D is the physical distance, and Δ is the risk level coefficient, where the risk coefficient Δ is 0.1 for level 1, 0.5 for level 2, and 1.0 for level 3.
[0013] Secondly, the present invention provides a fire evacuation method for small and medium-sized venues, comprising the following steps: Step 201: AI-based fire detection in small and medium-sized venues, using a lightweight AI model connected to cameras and linked devices for fire detection; Step 202: Dynamic evacuation route calculation. When a fire is detected in its very early stages, the monitoring camera captures and transmits the images to a lightweight AI model to further extract the smoke texture features of the images. At the same time, external temperature measurement points monitor the temperature changes in the area and collect 10 consecutive frames of images to detect the spread trend of the fire area and generate evacuation routes. Step 203: Intelligent indicator lights turn. The intelligent indicator lights set at critical path points control the direction of the indicator lights according to the evacuation path output by the lightweight AI model. Step 204: Voice-guided startup, further enabling voice alarms and mobile app information push notifications; Step 205: Indoor personnel are evacuated according to the evacuation routes; Step 206: Continuously monitor small and medium-sized venues. Once the fire is effectively controlled, return to step 201 and continue to monitor the environmental parameters within the venue using cameras and lightweight AI models.
[0014] Thirdly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the fire detection method for small and medium-sized locations as described in any of the first aspects.
[0015] Fourthly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the fire evacuation method for small and medium-sized locations as described in the second aspect.
[0016] Fifthly, the present invention provides a fire detection and emergency response device, comprising: One or more processors; Memory; and One or more computer programs, wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, wherein the processors, when executing the computer programs, implement the steps of the fire detection method for small and medium-sized locations as described in the third aspect.
[0017] Sixthly, the present invention provides a fire detection and emergency evacuation device, comprising: One or more processors; Memory; and One or more computer programs, wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, wherein the processors execute the computer programs to implement the steps of the fire evacuation method for small and medium-sized locations as described in the fourth aspect.
[0018] The lightweight fire detection and emergency evacuation device and method of this invention achieves accurate early-stage fire detection, rapid alarm information transmission, and intelligent evacuation route guidance through lightweight AI model optimization, multi-feature fusion recognition, dynamic evacuation path planning, and low-cost equipment linkage. This meets the core requirements of low cost, easy deployment, and high reliability for small and medium-sized venues. Attached Figure Description
[0019] Figure 1 Flowchart of fire detection methods for small and medium-sized venues.
[0020] Figure 2 Flowchart of fire evacuation methods for small and medium-sized venues.
[0021] Figure 3 : Schematic diagram of fire detection and emergency evacuation device. Detailed Implementation
[0022] To make the technical problems, technical solutions and beneficial effects of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.
[0023] Please refer to Figure 1 The flowchart of the fire detection method for small and medium-sized venues of the present invention includes the following steps: Step 101: Scene Modeling and Equipment Deployment in Small and Medium-Sized Venues. Considering the characteristics of small and medium-sized venues—limited space and simple layout—this step involves completing digital scene modeling and lightweight equipment deployment, specifically: First, scene topology modeling is performed using an image acquisition and simple annotation method. The internal layout of the venue (including the locations of passages, doors and windows, shelves, safety exits, and fire-fighting facilities) is captured by a mobile phone or ordinary camera. After importing the images into the system, a simplified scene topology map is automatically generated. Safety exit priorities are manually annotated, with main exits marked as level 1 and backup exits as level 2. Core parameters such as passage width and obstacle locations are also annotated. The accuracy of the passage width annotation reaches ±0.1m. The modeling time is ≤30 minutes. Secondly, lightweight equipment deployment is implemented: a low-cost solution of reusing existing equipment and adding new core equipment is adopted for equipment deployment: ① Existing surveillance cameras in the site are reused as image acquisition terminals and deployed at passageway intersections, densely populated equipment areas, and storage areas for flammable and explosive materials, with coverage without blind spots; ② New edge computing terminals are added. The edge computing terminals adopt ARM architecture, have at least 4GB of memory, and power consumption ≤15W. They are deployed next to the monitoring host to realize AI computing and data processing; ③ Linkage devices are deployed, including wireless emergency indicator lights, voice alarm modules, and personnel mobile terminal APPs.
[0024] Step 102: Training and Deploying a Lightweight AI Fire Detection Model. This involves building a lightweight AI model suitable for edge computing to achieve accurate multi-feature identification of fires in their very early stages. The specific steps include: First, a multi-source feature dataset was constructed: typical fire scene data of small and medium-sized venues were collected, including open flames and smoldering flames, dense smoke and faint smoke, and images of temperature anomalies. Combined with publicly available fire datasets, a dataset containing at least 100,000 samples was constructed, with each sample labeled with the "flame / smoke / normal" category and the corresponding temperature features. Secondly, a lightweight AI model design was implemented: Based on MobileNetV3 as the backbone network, the following optimizations were made: ① Model pruning: Redundant convolutional layers were removed, reducing the number of model parameters from 4.2M to 1.8M, and reducing the computational cost by 50%; ② Feature fusion module: A new "visual feature + temperature feature" fusion branch was added. Visual features extract the RGB channel difference of flame color and LBP features of smoke texture. Temperature features extract the mean temperature and mutation rate of the region. Feature weights are dynamically allocated through an attention mechanism; ③ Scene adaptation module: An illumination-adaptive convolutional kernel was introduced to automatically adjust image brightness and contrast, improving the recognition stability in low-light and high-light scenes. Model training and deployment are carried out again: the model is trained in the cloud using transfer learning, with a training accuracy of ≥99.2%. The model size is compressed to 0.8MB using INT8 quantization compression and deployed on an edge computing terminal with an inference speed of ≥30FPS to meet the real-time detection requirements.
[0025] Step 103: Real-time fire detection and risk assessment. The edge computing terminal processes camera image data in real time to complete fire identification and risk level determination, specifically including: First, image preprocessing and feature extraction are performed: After the camera transmits images with a frame rate of 15~30FPS to the edge terminal, Gaussian filtering and size normalization are performed as preprocessing, and then visual and temperature features are extracted through a lightweight model. Next, multi-feature fusion recognition is performed: the model outputs the recognition result after feature fusion. If the dual conditions of "visual feature matching fire type + temperature ≥ environmental baseline 20℃" are met, it is determined to be a fire signal; if only a single feature is abnormal, it is marked as "suspected risk" and three consecutive frames are verified. If three consecutive frames are abnormal, the real fire is confirmed. Finally, risk levels are classified: combining the proportion of the fire area in the image, the spread speed (obtained by calculating the expansion ratio of the fire area in 10 adjacent frames of images), and the distance from the safety exit, the risk is divided into 3 levels: Level 1 fire is a very early risk, with a fire area of <5% and no spread; Level 2 fire is a development stage, with a fire area of 5% to 20% and slow spread; Level 3 fire is a violent stage fire, with a fire area of >20% and rapid spread. Step 104: Dynamic evacuation route planning and updating. Based on the scene topology map and real-time fire data, an improved path planning algorithm is used to generate the optimal evacuation route, specifically including: First, real-time environmental data fusion is performed: fire location, risk level, and smoke spread direction are collected. The smoke spread direction is determined by real-time data such as the smoke movement trajectory in continuous frame images and the status of safety exits. The "dangerous area" and "passable area" markers in the scene topology map are then updated. Secondly, an improved route planning method is implemented: a risk level coefficient is introduced on the basis of the traditional route planning algorithm, and the route cost is calculated using the formula P = D × (1 + Δ), where P is the route cost, D is the physical distance, and Δ is the risk level coefficient, where the risk coefficient Δ is 0.1 for level 1, 0.5 for level 2, and 1.0 for level 3. At the same time, level 1 safety exits are given priority, and if the main exit is blocked by fire, it will automatically switch to level 2 backup exits to generate a dynamic evacuation route with "shortest distance + lowest risk". The route is adapted and updated again: In view of the dispersed distribution of people in small and medium-sized venues, the venue is divided into multiple evacuation sub-areas, and a dedicated route is generated for each sub-area. If the fire spreads and the risk of the original route increases to level 1 or above, the route is replanned and updated within 0.5 seconds.
[0026] Step 105: Multi-dimensional emergency alarms and evacuation guidance. Relying on deployed lightweight equipment, alarms and guidance are implemented through multiple dimensions including "visual + auditory + mobile terminals," forming a closed-loop response. Specifically, this includes: First, a tiered alarm triggering system is implemented: Level 1 risk triggers a local voice alarm and push notifications to the mobile app; Level 2 risk triggers a high-decibel voice alarm and flashing red emergency indicator lights; Level 3 risk triggers the highest level alarm, overlays evacuation instructions broadcast by all voice devices in the premises, and automatically uploads fire information to the 119 alarm platform. Secondly, dynamic guidance is implemented: ① Emergency indicator lights: The indicator direction is automatically switched according to the evacuation route, and the indicator lights near the danger zone display a red prohibition sign; ② Voice guidance: TTS technology is used to broadcast precise instructions in real time, such as "Please evacuate to the northeast safety exit and avoid the smoke and fire area on the left"; ③ The mobile APP pushes personalized evacuation route maps, marks the user's real-time location and target exit, and supports voice navigation.
[0027] Finally, feedback on the evacuation progress is provided: the evacuation status of people in each area is determined by recognizing human contours through camera images. If there are stranded people in a certain area, the alarm intensity and guidance frequency in that area are increased accordingly.
[0028] Step 106: Fire data analysis and detection model update. After the evacuation is completed, fire data is collected, analyzed, and transmitted to the lightweight AI fire detection model for training and updating to improve the accuracy of fire detection.
[0029] Further reference Figure 2 A flowchart of fire evacuation procedures for small and medium-sized venues, including the following steps: Step 201: AI-based fire detection in small and medium-sized venues, using a lightweight AI model connected to cameras and linked devices for fire detection; Step 202: Dynamic evacuation route calculation. When an early fire is detected, the monitoring camera captures and transmits the images to the lightweight AI model to further extract the smoke texture features of the images. At the same time, external temperature measurement points monitor the temperature changes in the area. Furthermore, 10 consecutive frames of images are used to detect the spread trend of the fire area and generate evacuation routes. Step 203: Intelligent indicator lights turn. The intelligent indicator lights set at critical path points control the direction of the indicator lights according to the evacuation path output by the lightweight AI model. Step 204: Voice-guided startup, further enabling voice alarms and mobile app information push notifications; Step 205: Indoor personnel are evacuated according to the evacuation routes; Step 206: Continuously monitor small and medium-sized venues. Once the fire is effectively controlled, return to the initial stage and continue to monitor the environmental parameters within the venue using cameras and lightweight AI models.
[0030] refer to Figure 3 A schematic diagram of a fire detection and emergency evacuation device. The fire detection and emergency evacuation device of the present invention further includes one or more memories 20 and one or more processors 30, wherein the one or more computer programs are stored in the memories 20 and configured to be executed by the one or more processors 30, and the processors 30 implement the steps of the fire detection and emergency response method when executing the computer programs.
[0031] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, which may include: read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.
[0032] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for fire detection in small and medium-sized venues, characterized in that, Includes the following steps: Step 101: Scene modeling and equipment deployment in small and medium-sized venues. Complete the digital modeling of the scene and the deployment of lightweight equipment to generate a scene topology map; Step 102: Training and deployment of lightweight AI fire detection model, building a lightweight AI model suitable for edge operation, and achieving accurate identification of multiple features of fires in the very early stage; Step 103: Real-time fire detection and risk assessment. The edge computing terminal processes camera image data in real time to complete fire identification and risk level determination. Step 104: Dynamic evacuation route planning and updating. Based on the scene topology map and real-time fire data, an improved path planning algorithm is used to generate the optimal evacuation route. Step 105: Multi-dimensional emergency alarm and evacuation guidance. Relying on the deployed lightweight equipment, alarms and guidance are achieved through multiple dimensions of vision, hearing and mobile terminals to form a closed-loop response. Step 106: Fire data analysis and detection model update. After evacuation, fire data is collected, analyzed, and transmitted to a lightweight AI fire detection model for training and updating to improve the accuracy of fire detection.
2. The fire detection method for small and medium-sized venues as described in claim 1, characterized in that, The lightweight AI fire detection model uses MobileNetV3 as its backbone network and has 1.8M model parameters.
3. The fire detection method for small and medium-sized venues as described in claim 2, characterized in that, The lightweight AI fire detection model further includes a feature fusion module, which extracts the RGB channel difference of flame color, LBP features of smoke texture, and regional temperature mean and mutation rate for feature fusion.
4. The fire detection method for small and medium-sized venues as described in claim 2, characterized in that, The lightweight AI fire detection model has a training accuracy of ≥99.2%, and its size is further compressed to 0.8MB through INT8 quantization compression.
5. The fire detection method for small and medium-sized venues as described in claim 1, characterized in that, The path planning uses the formula P=D×(1+Δ) to calculate the path cost, where P is the path cost, D is the physical distance, and Δ is the risk level coefficient, where the risk coefficient Δ is 0.1 for level 1, 0.5 for level 2, and 1.0 for level 3.
6. A fire evacuation method for small and medium-sized venues, characterized in that, Includes the following steps: Step 201: AI-based fire detection in small and medium-sized venues, using a lightweight AI model connected to cameras and linked devices for fire detection; Step 202: Dynamic evacuation route calculation. When a fire is detected in its very early stages, the monitoring camera captures and transmits the images to a lightweight AI model to further extract the smoke texture features of the images. At the same time, external temperature measurement points monitor the temperature changes in the area and collect 10 consecutive frames of images to detect the spread trend of the fire area and generate evacuation routes. Step 203: Intelligent indicator lights turn. The intelligent indicator lights set at critical path points control the direction of the indicator lights according to the evacuation path output by the lightweight AI model. Step 204: Voice-guided startup, further enabling voice alarms and mobile app information push notifications; Step 205: Indoor personnel are evacuated according to the evacuation routes; Step 206: Continuously monitor small and medium-sized venues. Once the fire is effectively controlled, return to step 201 and continue to monitor the environmental parameters within the venue using cameras and lightweight AI models.
7. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the fire detection method for small and medium-sized venues as described in any one of claims 1 to 5.
8. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the fire evacuation method for small and medium-sized venues as described in claim 6.
9. A fire detection and emergency response device, comprising: One or more processors; Memory; as well as One or more computer programs, wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, characterized in that the processors, when executing the computer programs, implement the steps of the fire detection method for small and medium-sized locations as described in claim 7.
10. A fire detection and emergency evacuation device, comprising: One or more processors; Memory; as well as One or more computer programs, wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, characterized in that the processors, when executing the computer programs, implement the steps of the fire evacuation method for small and medium-sized locations as described in claim 8.