CN-122020283-A - Multi-scene intelligent fire video detection method based on deep learning and dynamic difference

CN122020283ACN 122020283 ACN122020283 ACN 122020283ACN-122020283-A

Abstract

The invention relates to the technical field of artificial intelligence, in particular to a multi-scene intelligent fire video detection method based on deep learning and dynamic difference. The invention realizes all-round breakthrough in detection precision, environmental adaptability, cost benefit and scene expansibility through the core technical innovations of a dynamic differential and deep learning fusion dual verification mechanism, a multi-mode data collaborative decision, a lightweight modular design and the like.

Inventors

MA YUEFENG
ZHAO KAIJIE
WANG XIAOFENG
MAO YANZHE
WU YAO
HAN FANGJIE
YOU GUOQIANG
ZHAO HUAIPU
ZHENG JIANGUO
LIU KAI

Assignees

中国辐射防护研究院

Dates

Publication Date: 20260512
Application Date: 20251229

Claims (10)

1. The multi-scene intelligent fire video detection method based on deep learning and dynamic difference is characterized by being deployed in a layered architecture system, wherein the system comprises a perception layer, an edge calculation layer, a cloud collaboration layer and an application layer, and the method comprises the following steps: Step S1, acquiring real-time video stream and environmental sensor data of a monitoring area through a sensing layer; s2, carrying out dynamic differential processing on the real-time video stream at an edge computing layer to extract a moving target area; s3, adopting an optimized lightweight deep learning target detection model to identify a fire target in the moving target area, and obtaining a detection result; S4, based on a multi-mode fusion decision mechanism, carrying out fusion analysis on the dynamic differential processing result, the fire target identification result and the environmental sensor data to determine a fire event; and S5, executing hierarchical alarm and equipment linkage control through an application layer according to the judging result.
2. The method according to claim 1, wherein the dynamic differential processing in step S2 is specifically that a modified ViBe background modeling algorithm is adopted to extract a dynamic region and a space-time consistency check is added to distinguish fire spread motion from regular motion, and meanwhile, an adaptive frame sampling strategy based on scene activity is adopted to dynamically adjust the sampling frequency according to the differential energy between video frames.
3. The method of claim 1, wherein the lightweight deep learning object detection model in step S3 is an optimized YOLOv n model, the optimizing comprising introducing a coordinate attention mechanism in a model backbone network, improving a feature pyramid structure to enhance small object detection capability, and generating a special prior box using fire dataset clustering.
4. The method of claim 1, wherein the multi-mode fusion decision mechanism in step S4 specifically includes fusing motion features output by the dynamic differential module, detection confidence level output by the target detection model and environmental sensor readings, calculating a unified fire confidence value, and performing grading decision according to a preset threshold.
5. The method according to claim 1, wherein the step S5 of step alarm and device linkage control specifically comprises: if the early warning is judged, triggering a local audible and visual alarm to carry out early warning prompt, and pushing early warning information to a manager; And if the fire condition is determined to be confirmed, immediately triggering a fire alarm and executing linkage operation through the relay module, wherein the linkage operation comprises cutting off a non-fire-fighting power supply, starting the smoke discharging equipment and starting the fire extinguishing equipment after a preset delay.
6. The method of claim 1, further comprising a model iteration step of periodically transmitting the video frames and sensor data related to false alarm, early warning and fire confirmation back to the cloud collaboration layer by the edge calculation layer, updating the training data set by the cloud collaboration layer, performing incremental training on the deep learning target detection model, and transmitting the model after iterative optimization to the edge calculation layer for updating and deployment.
7. The method of claim 1, wherein an edge computing layer in the hierarchical architecture system is deployed in a local monitoring room for carrying core computing tasks of the dynamic differential processing, fire target identification, and multimodal fusion decision.
8. The method of claim 1, wherein the environmental sensor data comprises smoke sensor and temperature sensor data, and wherein the multimodal fusion decision mechanism utilizes sensor data to assist in verifying video detection results.
9. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the deep learning and dynamic differencing based multi-scenario intelligent fire video detection method according to any one of claims 1 to 8 when executing the program.
10. A computer readable storage medium having stored thereon a computer program, which when executed by a processor implements the deep learning and dynamic differencing based multi-scenario intelligent fire video detection method according to any of claims 1 to 8.

Description

Multi-scene intelligent fire video detection method based on deep learning and dynamic difference Technical Field The invention relates to the technical field of artificial intelligence, in particular to a multi-scene intelligent fire video detection method based on deep learning and dynamic difference. Background The traditional fire detection means often rely on manual judgment, fire sensor detection or monitoring means based on a single image processing technology, and have obvious defects in response speed and detection accuracy, such as response delay, easiness in environment (dust and chemical corrosion) interference, fixed installation position, limited detection space, inapplicability to large space, open space and other problems, and difficulty in meeting the actual requirements of early-stage rapid early warning of the fire. The invention is based on the fire video detection technology combining deep learning and dynamic difference, not only is helpful for realizing rapid and accurate identification in the early stage of fire, but also can obviously reduce false alarm and missing report rate in complex environment, and strives for precious time for emergency rescue, thereby guaranteeing the life and property safety of people. The system is integrated to the existing video monitoring system of a factory building and a warehouse, captures scene dynamic change characteristics through the built-in dynamic difference module, realizes quick detection of flame and smoke by combining the deep learning module, does not need newly-added hardware equipment, and meanwhile, links the traditional fire sensor to form a double verification mechanism, solves the problems of high false alarm rate and delayed initial fire response in dust and corrosive environments, and meets the real-time accurate monitoring requirement of fire. Potential application fields include industrial high-risk scenes (chemical plants, hazardous chemical warehouses) and public safety fields (shopping malls, subway hubs). A fire disaster video detecting and early warning method based on image multi-feature fusion is disclosed in Chinese patent CN110516609A, which comprises the steps of preprocessing after an image sequence of a video is acquired, extracting a foreground region, obtaining a detected candidate region, extracting static features and dynamic features from the candidate region, judging whether flame is contained or not by taking the static features and the dynamic features as inputs of an SVM classifier when flame is detected, logically combining, selecting and calculating judging results of the features to obtain whether the flame is contained or not when smoke is detected, judging fire disaster according to the growing trend of the flame or the smoke if the flame is detected, alarming fire disaster on a monitoring site when the fire disaster is judged to be formed, and otherwise, alarming fire disaster only in the background. The technology is combined with the existing monitoring system and is applied to places such as markets, storages and the like. Although a series of measures are taken in the prior art to improve the detection efficiency, on the premise of ensuring reliable detection, the delay still needs to be further improved, and the requirement of rapid early warning in the early stage of fire is difficult to meet. The technology mainly uses flame detection, has limited capability of detecting smoldering fire, is common in the early stage of fire, and can possibly lead to fire spreading if the smoldering fire cannot be detected in time. The technology lacks intelligent recognition capability under the condition of normal fire or controllable condition, and can not accurately early warn according to actual conditions. The main problems are represented by the following points: 1. The false alarm rate of smoldering flame/smoke images is high, and static flame images in billboards, posters and screens are easily misjudged as real fires, so that false alarm is caused. 2. The model is complex, and the storage space required for directly processing the high-definition video is large. 3. The recognition accuracy is not enough, and the detection performance of a model trained on an experimental data set only can be obviously reduced when the model is applied to actual conditions (light conditions, new scenes with large background environment differences). Disclosure of Invention The invention discloses a multi-scene intelligent fire video detection method based on deep learning and dynamic difference, which aims to solve the technical problems. The invention adopts the following technical scheme: The invention provides a multi-scene intelligent fire video detection method based on deep learning and dynamic difference, which is deployed in a hierarchical architecture system, wherein the system comprises a perception layer, an edge calculation layer, a cloud cooperation layer and an application layer, and the method comprises th