CN-121982641-A - Method and system for monitoring safety compliance of operators based on double-view fusion

CN121982641ACN 121982641 ACN121982641 ACN 121982641ACN-121982641-A

Abstract

The invention discloses a double-view fusion-based operator safety compliance monitoring method and system, which can be applied to a bottom blowing furnace operation scene of a copper factory. The method comprises the steps of collecting video of a working area through a double-view camera, and constructing a closed-loop monitoring system comprising a multi-element sensing module, a tracking detection cooperative system and a dynamic compliance judging engine. The cross-model feature multiplexing, the double-view target association, the active perception and re-detection, the state tracing and the correction are adopted to form a closed loop of detection, tracking, correction and judgment, the double-view depth coordination is realized, and the problems of safety state tracking fracture, risk level dynamic early warning missing and initial error accumulation caused by shielding are solved. The invention can realize continuous monitoring of the safety state and dynamic compliance judgment based on the risk area, has self-correction capability, and improves the reliability and instantaneity of the monitoring system.

Inventors

LI QI
ZHANG HUAN
ZHOU LIYONG
DU YONGXING
HU WEIJIAN

Assignees

内蒙古科技大学

Dates

Publication Date: 20260505
Application Date: 20260127

Claims (10)

1. A worker safety compliance monitoring method based on double-view fusion is characterized in that a double-view camera is used for synchronously collecting a synchronous video stream of an operation area, performing target detection and feature extraction, generating a tracking template through feature dimension adaptation, fusing multi-source features, integrating collected double-view information by means of an attention mechanism, constructing comprehensive tracking quality scores based on multi-dimension indexes, evaluating tracking quality, triggering local self-adaptive re-detection to update tracking track states and tracking template features, and finally combining time sequence consistency verification with dynamic risk area division and threshold adjustment to finish the process of outputting compliance judgment results and early warning signals.
2. The method for monitoring the safety compliance of operators based on double-view fusion according to claim 1, comprising the following steps: s1, acquiring a synchronous video stream of an operation area acquired by a double-view camera, performing target detection on each frame of image of the synchronous video stream, and outputting an initial detection result and a target feature vector of each preset detection target; s2, performing dimension adaptation on the target feature vector based on the learnable projection matrix, and generating and storing tracking template features corresponding to the track positions; s3, constructing a comprehensive feature matrix integrating the double-view template features, the search region features and the track query vectors, calculating attention scores of the track query vectors and the features of different view angles, and obtaining the fusion features by carrying out weighted summation on the attention score pair value matrix ; S4, constructing a comprehensive tracking quality score according to the feature matching score, the position stability score and the spatial visibility score; S5, when the comprehensive tracking quality score meets a preset re-detection condition, determining a local ROI (region of interest) based on the current track position, detecting the ROI to obtain a re-detection result, and updating the tracking track state and the tracking template characteristic according to the re-detection result; S6, calculating the time sequence consistency score of the state of the safety equipment through weighting of the sliding window, dynamically dividing a high-risk area by combining the state of the risk source and the position of the personnel, adjusting the judgment threshold value corresponding to each preset detection target, and outputting a compliance result and an early warning signal.
3. The method for monitoring the safety compliance of operators based on double-view fusion according to claim 2, further comprising, for stably tracking a trajectory of a preset number of frames, calculating a current detection frame With the initial detection frame Deviation metric value between Average feature matching confidence of current frame number When the preset correction condition is met, the current detection frame is used for replacing the initial detection frame and updating the tracking template characteristics.
4. The method for monitoring safety compliance of operators based on double-view fusion according to claim 3, wherein the expression of the initial detection result is: In the above-mentioned steps, Indicating the detection result of the ith target at the kth viewing angle, Representing the upper left corner coordinates of the bounding box, And Representing the width and height of the bounding box, The category of the object is indicated and, The degree of confidence is indicated and, Representing a high-dimensional feature vector of the target region.
5. The method for monitoring the safety compliance of operators based on double-view fusion according to claim 4, wherein the generation expression of the tracking template features is as follows: wherein Representing a matrix of projections that can be learned, Representing a high-dimensional feature vector of a kth object detected by the multi-element perception module at a t-th frame, The initial tracking template characteristic of the kth target is represented, T represents the tracking template, k represents the kth target, init represents initial generation, and the generated tracking template characteristic is bound with the track ID and stored in a template memory bank: ; Wherein track_id_k represents the track ID corresponding to the kth target, and memory_frames represents the storage structure of the template memory bank.
6. The method for monitoring the safety compliance of operators based on double-view fusion according to claim 5, wherein the comprehensive tracking quality score calculation formula is: , wherein, Representing the matching score of the double-view features, tracking the similarity mean value of the template and the double-view features for the current frame, Representing the position stability score, which is the average intersection ratio of the current frame and the target boundary frame of the previous three frames, Representing the spatial visibility score, which is the mesh duty cycle for which the target area is not occluded, 、、 Representing the weight coefficient.
7. The method for monitoring compliance with operator safety based on double view fusion according to claim 6, wherein the predetermined re-detection condition in step S5 is that And consecutive 3 frames satisfy And is also provided with Wherein , The expression of the local ROI area is as follows: wherein The upper left-hand abscissa, ordinate, The lower right-hand abscissa and the ordinate of the target bounding box of the t frame are represented; A pixel.
8. The method for monitoring safety compliance of operators based on double-view fusion according to claim 7, wherein the preset correction conditions are: And is also provided with Wherein As a measure of the deviation is calculated, For the detection frame of the kth object at the t frame, Is an initial detection box for the target, wherein, Represents the confidence of the average feature matching of near M frames, M represents the number of frames of a long time sequence window, Representing the dual view feature matching score for the i-th frame.
9. The method for monitoring the safety compliance of operators based on double-view fusion according to claim 8, wherein the calculation formula of the time sequence consistency score is as follows: wherein c represents a category, t represents a frame number, W represents a sliding window size, The weight of the kth frame is represented, and the weight is reset to 0 in the shielding state; representing class c timing consistency scores for the ith target at the t-th frame, Representing class c single frame detection scores for the ith target at the kth frame.
10. An operator safety compliance monitoring system based on double-view fusion, comprising: The double-view video acquisition module consists of two high-definition cameras with different view angles and is used for synchronously acquiring video streams of an operation area; The multi-element perception module is used for integrating the target detection model, receiving the double-view video frame, outputting an initial detection result and a target feature vector of each preset detection target and providing a data basis for the subsequent module; the tracking detection cooperative system comprises a cross-model feature multiplexing unit, a double-view target association unit, an active sensing and re-detection unit and a state tracing and correcting unit, and is used for realizing cross-view stable track tracking and continuous updating; the dynamic compliance judging engine comprises a time sequence consistency checking unit, a risk area dividing unit and a compliance rule judging unit, and is used for fusing the track and the risk source information and outputting a dynamic compliance result and an early warning signal; and the early warning output module is used for receiving the compliance judgment result and outputting an early warning signal.

Description

Method and system for monitoring safety compliance of operators based on double-view fusion Technical Field The invention particularly relates to an operator safety compliance monitoring method and system based on double-view fusion, and belongs to the technical field of industrial safety monitoring. Background In an industrial safety high-risk operation scenario, operators must be strictly protected with equipment. The automatic safety monitoring technology based on computer vision is a key means for guaranteeing the safety production of the scene, and the existing visual automatic safety monitoring method is mainly divided into two types, namely, one type is based on target detection of images, and the other type is based on multi-target tracking of single-view video; The image-based target detection method mainly analyzes a single-frame image through a deep learning model, and identifies personnel and safety equipment. Although the method can identify the safety equipment in a single frame, the method cannot judge the continuity of the wearing state, and can easily generate missing report and false report in copper factory workshops with frequent personnel shielding and gesture change, and cannot meet the continuous supervision requirement. In the prior art, a multi-target tracking method based on single-view video mainly generates a motion track through a single camera associated continuous frame detection frame in a bottom blowing furnace operation area of a copper factory, but the single-view camera has an unavoidable monitoring blind area, and the monitoring effects of a front view and a side view have complementarity as shown in fig. 1 (b) and 1 (c), but even if the monitoring effects have complementarity, double-view information cannot be effectively fused, so that the special technical problem of three copper factory safety supervision is caused. Firstly, when key safety equipment is shielded under a certain view angle due to personnel operation, the system cannot effectively utilize the complementary information of another view angle, so that the safety state tracking is broken, and continuous supervision cannot be realized. The existing method cannot fuse the relationship between accurate positioning personnel and a high Wen Fengxian source through double visual angles, cannot realize dynamic monitoring and early warning based on real risk levels, and finally in a double visual angle system, errors of an initial detection frame of any visual angle can be continuously accumulated in a tracking chain, and the system lacks a cross visual angle cooperative checking mechanism to correct the initial errors, so that misjudgment can influence the reliability of a final conclusion. Firstly, the security state tracking is broken, when a key security device is blocked at a certain view angle, the system cannot utilize the complementary information of the other view angle, so that the security state tracking is interrupted, continuous supervision cannot be realized, secondly, the dynamic risk early warning is lost, the existing method cannot fuse the relationship between a precisely positioned person and a high Wen Fengxian source through the double view angles, dynamic monitoring and early warning cannot be realized based on the real risk level, the initial detection error of any view angle in the double view angle system can be continuously accumulated in a tracking chain, and the error is corrected due to the lack of a cross view angle collaborative verification mechanism, so that erroneous judgment is generated. Therefore, a method for solving the above technical problems is needed. Disclosure of Invention The invention aims to provide an operator safety compliance monitoring method based on double-view fusion, which is used for synchronously acquiring synchronous video streams of an operation area through a double-view camera, carrying out target detection and feature extraction, integrating acquired double-view information by means of a attention mechanism after feature dimension adaptation to generate a tracking template, integrating multi-source features, constructing comprehensive tracking quality scores based on multi-dimension indexes, evaluating tracking quality, triggering local self-adaptive weight detection to update tracking track states and tracking template features, and finally completing the process of outputting compliance judgment results and early warning signals by combining time sequence consistency verification with dynamic risk area division and threshold adjustment. Preferably, the method comprises the following steps: s1, acquiring a synchronous video stream of an operation area acquired by a double-view camera, performing target detection on each frame of image of the synchronous video stream, and outputting an initial detection result and a target feature vector of each preset detection target; s2, performing dimension adaptation on the target feature vector based on the learn