CN-122024394-A - Fire smoke visual identification and intelligent alarm method and system

CN122024394ACN 122024394 ACN122024394 ACN 122024394ACN-122024394-A

Abstract

The invention discloses a fire smoke visual identification and intelligent alarm method and system, which relate to the technical field of fire control monitoring and acquire and preprocess video streams to obtain standardized frame data, extract characteristics through a modified YOLOv model and identify smoke candidate areas to exclude non-smoke interference, extract four-dimensional refined characteristics from the candidate areas, verify whether fire smoke exists through an SVM fire identification model, quantitatively judge fire levels and match a grading alarm threshold, trigger local alarm when the threshold is reached, upload a fire platform and link emergency equipment. The invention combines deep learning and multi-feature fusion technology, improves the accuracy and real-time performance of smoke identification, realizes the intelligent linkage of fire grading alarm and emergency equipment, and is suitable for various indoor and outdoor large-scale monitoring scenes.

Inventors

PAN WEI
WU HUI
PAN CHUN
WU DANDAN
ZHENG SHANGCHAO

Assignees

浙江海康威名科技有限公司

Dates

Publication Date: 20260512
Application Date: 20260303

Claims (9)

1. The fire smoke visual identification and intelligent alarm method is characterized by comprising the following steps of: S1, acquiring field video stream data to be monitored in real time, synchronously extracting and preprocessing frames of the video stream, filtering noise, illumination distortion and lens interference, and obtaining standardized visual frame data; s2, carrying out feature extraction on standardized visual frame data based on a deep learning model, identifying candidate areas which accord with smoke morphology, gray level distribution and dynamic diffusion features in the frame, eliminating non-smoke interference targets, and outputting smoke candidate area images; S3, carrying out refined feature analysis on the smoke candidate region image, extracting texture complexity, edge ambiguity, motion trail and infrared absorption coefficient features of smoke, establishing a fire disaster identification model to judge whether the candidate region is fire disaster smoke, and returning to the step S1 for continuous acquisition and monitoring if the candidate region is non-fire disaster smoke; S4, quantitatively judging fire grades based on the fire smoke characteristic parameters which pass verification, and correspondingly matching with preset grading alarm thresholds; S5, when the fire grade reaches the corresponding alarm threshold, triggering a local audible and visual alarm device, and uploading the fire position, grade, real-time picture and characteristic data to a fire monitoring platform, and starting emergency lighting, smoke discharging equipment and an area isolation device in a linkage way to complete intelligent alarm response.
2. The fire smoke visual identification and intelligent alarm method according to claim 1 is characterized in that an improved YOLOv model is adopted as the deep learning model, CSPDARKNET is used as a backbone network for feature extraction, a PAN-FPN structure is adopted as a neck for feature fusion, a detection head is adopted as a head for candidate region prediction, and a SE-Net attention mechanism is added into a C3 module of the backbone network to enhance the extraction capacity of the model on smoke features and inhibit interference of irrelevant features.
3. The fire smoke visual recognition and intelligent alarm method according to claim 2, wherein the specific principle of the SE-Net attention mechanism is as follows: carrying out global average pooling on the feature images output by the backbone network to obtain channel-level feature vectors; carrying out nonlinear transformation on the channel-level feature vector through the full-connection layer to obtain the channel attention weight; and carrying out channel-by-channel weighting on the channel attention weight and the feature map output by the backbone network to obtain an enhanced feature map.
4. The method for visually identifying and intelligently alarming fire smoke according to claim 1 is characterized in that the specific mode of eliminating non-smoke interference targets is that an interference target feature library is established, the interference target feature library comprises texture features of static background, granularity features of raised dust and gray threshold features of water vapor, the extracted candidate region features are compared with the interference target feature library, and if the similarity is lower than a preset threshold, the non-smoke interference targets are judged to be eliminated.
5. The fire smoke visual identification and intelligent alarm method is characterized in that a multi-feature fusion algorithm is adopted in the refined feature analysis, texture complexity is quantified by extracting energy, entropy and contrast parameters through a gray level co-occurrence matrix, edge ambiguity is judged by extracting an edge pixel duty ratio and an edge gradient value through a Canny edge detection algorithm, a motion track is obtained by extracting displacement vectors of smoke pixels between frames through an optical flow method and fitting the displacement vectors, an infrared absorption coefficient feature is synchronously acquired through an infrared imaging module, and the infrared absorption coefficient feature is calculated by combining the absorption characteristic of smoke to infrared light with specific wavelength.
6. The method for visually identifying and intelligently alarming fire smoke according to claim 1 is characterized in that the fire identification model is a classification model based on a support vector machine, the model takes texture complexity, edge ambiguity, motion trail and infrared absorption coefficient of smoke as input characteristics, a candidate area is output after training to be a judging result of fire smoke or non-fire smoke, and the training process adopts a fire smoke sample set and a non-smoke interference sample set to conduct supervision training until the model identification accuracy reaches a preset threshold.
7. The method for visually recognizing and intelligently alarming fire smoke according to claim 1, wherein the recognition process of the fire recognition model is specifically as follows: normalizing the extracted four-dimensional refined features to construct feature vectors; The method comprises the steps of adopting a radial basis function as a kernel function of an SVM, adopting a training set to train an SVM model, optimizing model parameters through a cross validation method, finding an optimal classification hyperplane, setting a classification decision function, judging fire smoke when the classification decision function is output to be 1, judging non-fire smoke when the classification decision function is output to be-1, and stopping training until the identification accuracy of the model on the validation set reaches a preset threshold value to obtain a trained fire identification model; The feature vector is input into a trained fire disaster recognition model to obtain a judging result, if the feature vector is judged to be non-fire disaster smoke, the step S1 is returned to continuously collect the on-site video stream data to be monitored for cyclic monitoring, and if the feature vector is judged to be fire disaster smoke, the step S4 is entered to quantify fire disaster grades.
8. The method for visually recognizing and intelligently alarming fire smoke according to claim 1, wherein the fire grade is classified into three grades of light, medium and heavy, the quantitative judgment indexes comprise the proportion of the smoke diffusion area to the area of the monitoring area, the smoke concentration and the diffusion rate, the preset grading alarm threshold corresponds to the three fire grades, the light alarm threshold, the medium alarm threshold and the heavy alarm threshold are respectively set, and each threshold corresponds to a group of smoke characteristic parameter ranges.
9. A fire smoke visual identification and intelligent alarm system, characterized in that a fire smoke visual identification and intelligent alarm method according to any one of claims 1-8 is applied, comprising: the data acquisition module acquires the on-site video stream data to be monitored in real time, synchronously performs frame extraction and preprocessing on the video stream, and filters noise, illumination distortion and lens interference to obtain standardized visual frame data; The feature extraction module is used for carrying out feature extraction on the standardized visual frame data based on the deep learning model, identifying candidate areas which accord with smoke morphology, gray level distribution and dynamic diffusion features in the frame, eliminating non-smoke interference targets and outputting smoke candidate area images; The fire disaster identification module is used for carrying out refined feature analysis on the smoke candidate area image, extracting texture complexity, edge ambiguity, motion trail and infrared absorption coefficient characteristics of smoke, establishing a fire disaster identification model to judge whether the candidate area is fire disaster smoke, and returning to the data acquisition module for continuous acquisition and monitoring if the candidate area is non-fire disaster smoke; The fire grade quantifying module quantifies and judges the fire grade based on the fire smoke characteristic parameters passing verification, and correspondingly matches a preset grading alarm threshold; and when the fire grade reaches the corresponding alarm threshold, the intelligent alarm module triggers the local audible and visual alarm device, and simultaneously uploads the fire position, grade, real-time picture and characteristic data to the fire monitoring platform, and the emergency lighting device, the smoke discharging device and the regional isolation device are started in a linkage way to complete intelligent alarm response.

Description

Fire smoke visual identification and intelligent alarm method and system Technical Field The invention relates to the technical field of fire control monitoring, in particular to a fire smoke visual identification and intelligent alarm method and system. Background At present, a fire disaster is a disaster with extremely high hazard, and a large amount of smoke is generated at the initial stage of the fire disaster, so that the fire disaster smoke is recognized in time and an alarm is given out, thereby being capable of striving for valuable time for fire disaster suppression and reducing casualties and property loss. The existing fire smoke alarm methods are mainly divided into two types, namely an alarm method based on a sensor, such as a smoke sensor, a temperature sensor and the like, the method is simple in structure and low in cost, but has the problems of limited detection range, easiness in environmental interference (such as dust emission, water vapor, temperature change), alarm lag and the like, and is difficult to meet the requirements of large indoor and outdoor monitoring scenes, and an alarm method based on visual recognition, wherein video data are collected through a camera and smoke is recognized by combining an image recognition technology, but the existing visual recognition method mainly adopts a traditional image processing algorithm, the extraction of smoke characteristics is not comprehensive enough, non-smoke targets such as dust emission, water vapor and shadow are easily misjudged as smoke, the recognition accuracy is low, and a fire grading judgment and emergency equipment linkage mechanism is lacked, so that intelligent alarm response cannot be realized. Therefore, how to improve the accuracy and the instantaneity of smoke identification through deep learning and multi-feature fusion technology and realize the linkage of fire grading alarm and emergency equipment is a problem to be solved by the technicians in the field. Disclosure of Invention In view of the above, the invention provides a method and a system for visual identification and intelligent alarm of fire smoke, which solve the problems in the background technology. In order to achieve the above purpose, the present invention adopts the following technical scheme: a fire smoke visual identification and intelligent alarm method comprises the following steps: S1, acquiring field video stream data to be monitored in real time, synchronously extracting and preprocessing frames of the video stream, filtering noise, illumination distortion and lens interference, and obtaining standardized visual frame data; s2, carrying out feature extraction on standardized visual frame data based on a deep learning model, identifying candidate areas which accord with smoke morphology, gray level distribution and dynamic diffusion features in the frame, eliminating non-smoke interference targets, and outputting smoke candidate area images; S3, carrying out refined feature analysis on the smoke candidate region image, extracting texture complexity, edge ambiguity, motion trail and infrared absorption coefficient features of smoke, establishing a fire disaster identification model to judge whether the candidate region is fire disaster smoke, and returning to the step S1 for continuous acquisition and monitoring if the candidate region is non-fire disaster smoke; S4, quantitatively judging fire grades based on the fire smoke characteristic parameters which pass verification, and correspondingly matching with preset grading alarm thresholds; S5, when the fire grade reaches the corresponding alarm threshold, triggering a local audible and visual alarm device, and uploading the fire position, grade, real-time picture and characteristic data to a fire monitoring platform, and starting emergency lighting, smoke discharging equipment and an area isolation device in a linkage way to complete intelligent alarm response. Optionally, the deep learning model adopts an improved YOLOv model, the model takes CSPDARKNET as a backbone network and is responsible for feature extraction, the neck adopts a PAN-FPN structure and is responsible for feature fusion, the head adopts a detection head and is responsible for candidate region prediction, and the SE-Net attention mechanism is added into a C3 module of the backbone network, so that the extraction capacity of the model on smoke features is enhanced, and the interference of irrelevant features is inhibited. Optionally, the specific principle of the SE-Net attention mechanism is as follows: carrying out global average pooling on the feature images output by the backbone network to obtain channel-level feature vectors; carrying out nonlinear transformation on the channel-level feature vector through the full-connection layer to obtain the channel attention weight; And carrying out channel-by-channel weighting on the channel attention weight and the feature map output by the backbone network to obtain an enhanced