Search

CN-121982648-A - Chemical industry park personnel action monitoring system that violating rules based on pattern recognition

CN121982648ACN 121982648 ACN121982648 ACN 121982648ACN-121982648-A

Abstract

The invention discloses a chemical industry park personnel illegal action monitoring system based on pattern recognition, which comprises the steps of collecting continuous video streams of an operation area, generating a time stamp video frame, calculating brightness statistics to form an illumination fluctuation sequence, detecting personnel candidate areas to obtain a boundary frame sequence, inputting the time stamp video frame and the boundary frame into a SAM 2 streaming memory video segmentation process to obtain a foreground mask, generating visibility confidence and shielding event marks by consistency indexes, encoding the foreground mask, the visibility confidence and the shielding event marks into shielding perception guide conditions, inputting LightenDiffusion to decompose and denoise the foreground mask to generate an enhanced video frame, constructing action fragments based on the enhanced video frame and the foreground mask, extracting appearance, geometry and motion characteristics to form action characterization and finishing illegal action discrimination and event recording. The invention reduces false alarm missing report under complex shielding and illumination change conditions.

Inventors

  • YI CHENG
  • YANG MEI

Assignees

  • 贵州兴园云信科技有限公司

Dates

Publication Date
20260505
Application Date
20260130

Claims (9)

  1. 1. Chemical industry park personnel action monitoring system that violating regulations based on mode discernment, characterized by, include the following module: The video acquisition module acquires continuous video streams of a chemical industry park operation area, adds time stamps to video frames and arranges the video frames in time sequence to form a video frame sequence, calculates brightness statistics frame by frame to form an illumination fluctuation sequence; the candidate detection module is used for executing personnel candidate area detection on the video frame sequence, generating a personnel candidate area boundary frame set of each frame, and forming a personnel candidate area sequence according to time sequence; The memory segmentation module inputs the video frame sequence and the personnel candidate region sequence into a SAM 2 streaming memory video segmentation process to generate a personnel foreground mask sequence, calculates an area change rate, a boundary consistency metric value and a cross-frame matching confidence coefficient based on the foreground mask sequence, fuses the two to generate a visibility confidence coefficient sequence, and generates an occlusion event mark sequence; The shielding guide module encodes the foreground mask sequence, the visibility confidence coefficient sequence and the shielding event mark sequence into shielding perception guide conditions, and accesses LightenDiffusion the shielding perception guide conditions into a potential space Retinex decomposition flow to execute foreground gating constraint on the reflection component generation path and the illumination component generation path; the diffusion updating module is used for accessing LightenDiffusion the shielding perception guiding conditions into a LightenDiffusion diffusion denoising iterative process, calling a streaming memory state retrieval historic foreground structure prior in a frame segment corresponding to a shielding event mark, and updating a front Jing Qian variable; The enhancement generation module is used for processing the video frame sequence by utilizing a potential space Retinex decomposition flow and a diffusion denoising iteration flow to generate an enhancement video frame sequence; the action construction module is used for constructing personnel action fragments based on the enhanced video frame sequence and the foreground mask sequence, extracting appearance characteristics, geometric characteristics and motion characteristics from the action fragments and fusing the appearance characteristics, the geometric characteristics and the motion characteristics to form an action characterization sequence; And the violation judging module inputs the action characterization sequence into a violation action judging process, generates action category distribution and violation confidence distribution, generates an alarm threshold based on the illumination fluctuation sequence and the visibility confidence sequence, and generates a violation event record when the violation confidence exceeds the alarm threshold.
  2. 2. The pattern recognition-based chemical industry park personnel violation monitoring system of claim 1, wherein the video acquisition module specifically comprises: collecting continuous video streams of a chemical industry park operation area and outputting continuous video frames; generating time stamps for continuous video frames frame by frame and arranging the time stamps in sequence to obtain a video frame sequence; And calculating brightness statistics on the video frame sequence frame by frame and arranging the brightness statistics in time sequence to obtain an illumination fluctuation sequence.
  3. 3. The pattern recognition-based chemical industry park personnel violation monitoring system of claim 1, wherein the candidate detection module specifically comprises: performing pixel value range uniformization processing on the video frames to obtain a preprocessing video frame sequence; generating a candidate search area in each video frame of the preprocessed video frame sequence, wherein the candidate search area is determined by the full-width area of the video frame; performing personnel candidate region detection on the preprocessed video frame sequence in the candidate search region to generate a personnel candidate region boundary box set, wherein the personnel candidate region boundary box set consists of personnel candidate region boundary boxes; Performing boundary frame coordinate normalization processing on the personnel candidate region boundary frame set and forming a boundary frame coordinate set, wherein the boundary frame coordinate set is arranged according to the sequence of video frame time stamps; performing bounding box overlap suppression processing on the bounding box coordinate sets, and outputting a reserved bounding box set of each video frame; And arranging the reserved boundary box sets of the video frames in a time stamp sequence to form a personnel candidate region sequence.
  4. 4. The pattern recognition-based chemical industry park personnel violation monitoring system of claim 1, wherein the memory segmentation module comprises: Establishing a frame-level corresponding relation between a video frame and a person candidate area boundary frame set according to the video frame time stamp to obtain a frame-level corresponding relation set; extracting candidate region pixel data in a limiting range of a personnel candidate region boundary box of each video frame according to the frame-level corresponding relation set, and establishing a candidate region input set according to a personnel candidate region boundary box index to obtain a candidate region input sequence arranged according to a time stamp sequence; Inputting a candidate region input set corresponding to the first position of a time stamp sequence in a candidate region input sequence into a SAM 2 streaming memory video segmentation process, generating a personnel foreground mask set, and writing the personnel foreground mask set into a streaming memory state to obtain an initial streaming memory state; The method comprises the steps of calling an initial streaming memory state as streaming memory state input of a SAM 2 streaming memory video segmentation process for a candidate region input set corresponding to each timestamp except for the first bit of a timestamp sequence in a candidate region input sequence, and generating a personnel foreground mask set corresponding to the timestamp; Performing index consistency verification on the personnel foreground mask set of the adjacent timestamp and the personnel candidate region boundary box set of the adjacent timestamp, and performing mask index rearrangement on the personnel foreground mask set according to the index consistency verification result to obtain a personnel foreground mask set aligned according to the personnel candidate region boundary box index; Writing a personnel foreground mask set aligned according to personnel candidate region boundary frame indexes and a candidate region input set corresponding to a timestamp into a streaming memory state, updating the streaming memory state and inputting a SAM 2 streaming memory video segmentation flow for the next timestamp to obtain a streaming memory state sequence updated by time recursion; and arranging the personnel foreground mask sets corresponding to the time stamps in sequence according to the time stamps to form a personnel foreground mask sequence.
  5. 5. The pattern recognition-based chemical industry park personnel violation monitoring system of claim 1, wherein the occlusion guidance module comprises: establishing an adjacent time stamp pairing set based on the personnel foreground mask sequence and according to the time stamp to obtain the adjacent time stamp pairing set; Calculating personnel foreground mask areas of each pair of personnel foreground masks in the adjacent timestamp pairing set and calculating the area change rate of the personnel foreground masks of the adjacent timestamps to obtain an area change rate sequence arranged according to the timestamp sequence; calculating the area of an overlapping area and the area of a union area for each pair of personnel foreground masks in the adjacent timestamp pairing sets, and calculating boundary consistency metric values according to the area of the overlapping area and the area of the union area to obtain a boundary consistency metric value sequence arranged according to the timestamp sequence; calculating the area of an overlapping area for each pair of personnel foreground masks in the adjacent timestamp pairing set, calculating the area of the personnel foreground mask of the current timestamp, and calculating the cross-frame matching confidence according to the area of the overlapping area and the area of the personnel foreground mask of the current timestamp, so as to obtain a cross-frame matching confidence sequence arranged in a timestamp sequence; Performing normalization processing on the area change rate sequence, the boundary consistency measurement value sequence and the cross-frame matching confidence coefficient sequence, and performing arithmetic average on normalization processing results to obtain a visibility confidence coefficient sequence arranged according to the time stamp sequence; And executing shielding event marking on the time stamp according to the visibility confidence sequence to obtain a shielding event marking sequence.
  6. 6. The pattern recognition-based chemical industry park personnel violation monitoring system of claim 1, wherein the diffusion updating module comprises: Based on the personnel foreground mask sequence, the visibility confidence coefficient sequence and the shielding event marking sequence, establishing ternary corresponding relations among the personnel foreground mask, the visibility confidence coefficient and the shielding event marking according to the time stamp, and obtaining a ternary corresponding relation set; Performing size alignment processing on the personnel foreground mask corresponding to the time stamp in the ternary corresponding relation set to enable the personnel foreground mask size to be consistent with the video frame size corresponding to the time stamp, and obtaining a size alignment personnel foreground mask sequence; Performing numerical mapping processing on the visibility confidence coefficient corresponding to the timestamp and the shielding event mark in the ternary corresponding relation set, and establishing association between a numerical mapping result and a size alignment personnel foreground mask according to the timestamp to obtain a guide parameter set corresponding to the timestamp; Performing coding processing according to the guide parameter set corresponding to the time stamp and the size alignment personnel foreground mask sequence to form shielding perception guide conditions; accessing LightenDiffusion the shielding perception guiding conditions into a reflection component generation path input of a latent space Retinex decomposition flow, and executing foreground gating constraint on the reflection component generation path according to a size alignment personnel foreground mask to obtain a latent space reflection component representation sequence; And accessing LightenDiffusion the shielding perception guiding conditions into an illumination component generation path input of a potential space Retinex decomposition flow, and executing foreground gating constraint on the illumination component generation path according to a size alignment personnel foreground mask to obtain a potential space illumination component representation sequence.
  7. 7. The pattern recognition-based chemical industry park personnel violation monitoring system of claim 1, wherein the enhancement generation module comprises: Based on the video frame sequence, the shielding perception guiding condition, the latent space reflection component representation sequence and the latent space illumination component representation sequence, building quaternary corresponding relations among the video frame, the shielding perception guiding condition, the latent space reflection component representation and the latent space illumination component representation according to the time stamp, and obtaining a quaternary corresponding relation set; Inputting LightenDiffusion a latent space Retinex decomposition process to the video frame corresponding to the timestamp in the quaternary corresponding relation set, and taking the shielding sensing guide condition corresponding to the timestamp and the latent space reflection component representation and the latent space illumination component representation as the input conditions of the decomposition process to obtain an updating result of the latent space reflection component representation and the latent space illumination component representation corresponding to the timestamp; The latent space reflection component representation updating result corresponding to the time stamp and the latent space illumination component representation updating result are input LightenDiffusion to diffuse and denoise the iterative flow with the shielding perception guiding condition corresponding to the time stamp, and the front Jing Qian variable updating result corresponding to the time stamp is generated; executing video frame generation processing on the front Jing Qian variable updating result corresponding to the time stamp to obtain an enhanced video frame corresponding to the time stamp; And arranging the enhanced video frames corresponding to the time stamps in the time stamp sequence to generate an enhanced video frame sequence.
  8. 8. The pattern recognition-based chemical industry park personnel violation action monitoring system of claim 1, wherein the action construction module comprises: Based on the enhanced video frame sequence and the personnel foreground mask sequence, establishing a frame-level corresponding relation between the enhanced video frame and the personnel foreground mask according to the time stamp to obtain a frame-level corresponding relation set; performing mask boundary box calculation on the personnel foreground mask corresponding to the time stamp based on the frame-level corresponding relation set to obtain a personnel foreground mask boundary box set, and cutting personnel image fragments from the enhanced video frame according to the personnel foreground mask boundary box set to obtain a personnel image fragment sequence arranged according to the time stamp sequence; performing sliding window segmentation on the personnel image fragment sequence according to the fixed frame length to obtain a personnel action fragment set; Calculating appearance characteristics of each personnel action segment in the personnel action segment set, wherein the appearance characteristics consist of a color histogram and texture statistics of personnel image segments, and an appearance characteristic sequence is obtained; calculating geometric features for each human motion segment in the human motion segment set, wherein the geometric features consist of center coordinates, width, height and length-width ratio of a human foreground mask boundary frame, and a geometric feature sequence is obtained; Calculating motion characteristics of each personnel action segment in the personnel action segment set, wherein the motion characteristics consist of the center coordinate difference of the personnel foreground mask boundary frames of adjacent time stamps and the personnel foreground mask area difference, and a motion characteristic sequence is obtained; and performing feature stitching on the appearance feature sequence, the geometric feature sequence and the motion feature sequence to obtain an action characterization sequence.
  9. 9. The pattern recognition-based chemical industry park personnel violation monitoring system of claim 1, wherein the violation discrimination module comprises: establishing a motion characterization vector set based on the motion characterization sequence and a timestamp, and arranging the motion characterization vector set according to the index of the personnel motion segment to obtain a motion discrimination input sequence; vector dimension alignment processing and numerical normalization processing are carried out on the action discrimination input sequence, and a normalized action discrimination input sequence is obtained; Performing time sequence feature mapping calculation on the normalized action discrimination input sequence according to time sequence, and performing category mapping calculation on the time sequence feature mapping calculation result to obtain action category distribution; Extracting the corresponding probability value of the illegal action category based on the action category distribution and arranging the probability value according to the time stamp to obtain illegal confidence distribution; establishing a numerical corresponding relation based on the illumination fluctuation sequence and the visibility confidence sequence according to the time stamp, performing normalization processing on the numerical corresponding relation, and taking arithmetic average to obtain an alarm threshold; and comparing the violation confidence distribution with the alarm threshold value according to the time stamp, and writing the violation event record to obtain the violation event record arranged according to the time stamp.

Description

Chemical industry park personnel action monitoring system that violating rules based on pattern recognition Technical Field The invention relates to the technical field of intelligent monitoring of industrial safety video and computer vision pattern recognition, in particular to a chemical industry park personnel illegal action monitoring system based on pattern recognition. Background The chemical industry garden operation area has the production characteristics of high temperature and high pressure, inflammable and explosive, poisonous and harmful substances coexistence, and personnel enter the device area, the storage and transportation area and the loading and unloading area, do not wear protective articles according to the regulation, break into the forbidden area through the line, climb and cross over the fence, carry out illegal actions such as unlicensed operation in the limited area, and the like, and all the safety risks can be possibly caused. Therefore, parks often deploy video monitoring and edge computing devices to monitor and alert personnel activities online to support security management and emergency handling. The existing monitoring scheme of the illegal actions of the personnel mostly adopts a video-based mode recognition route, firstly, personnel target detection is carried out on video frames to obtain personnel candidate area boundary boxes, and then, the characterization and classification discrimination are carried out on the actions of the personnel. In order to improve the motion discrimination stability, a part of schemes are introduced into multi-frame association and target tracking, adjacent frame personnel candidate areas are matched to form motion fragments, a part of schemes are introduced into foreground segmentation or human body contour extraction, personnel foreground is separated from complex background and then appearance, geometry and motion characteristics are extracted, a part of schemes are introduced into image enhancement and illumination correction, interference of illumination fluctuation on the motion characteristics is restrained through decomposition of reflection components and illumination components, and an alarm threshold is set according to environmental brightness change in a subsequent discrimination stage. Common equipment shielding, pipe gallery cross shielding, personnel mutual shielding, reflection glare and other conditions on the site of the chemical industry park lead to the situation that personnel candidate area boundary boxes are cut off, offset and overlapped, and foreground segmentation results easily produce fluctuation in the aspect of cross-frame consistency, so that action segment construction is broken or mismatched. Meanwhile, the brightness distribution of the video frames can be changed by day-night alternation, local light filling, steam smoke and shadow change, and the comparability of appearance characteristics and motion characteristics is affected. When the shielding and illumination fluctuation superposition occurs, abnormal jump is easy to occur in the action category distribution and the illegal confidence distribution, and the alarm threshold generation and comparison process is difficult to stably describe the illegal judgment boundary, so that the illegal actions of personnel in the chemical industry park are judged in real time, and the false alarm rate is high under the condition of complex shielding illumination change. Therefore, how to provide a chemical industry park personnel illegal action monitoring system based on pattern recognition is a problem that needs to be solved by the person skilled in the art. Disclosure of Invention The invention aims to provide a chemical industry park personnel illegal action monitoring system based on pattern recognition, and provides the technical scheme based on continuous video stream acquisition and illumination fluctuation modeling, personnel candidate area detection, video foreground segmentation and visibility confidence calculation based on stream memory, shielding perception guiding condition coding, foreground gating constraint of latent space reflection components and illumination component generation paths, front Jing Qian variable updating combined with stream memory state retrieval in diffusion denoising iteration, enhanced video frame sequence generation, action fragment construction and apparent geometric motion feature fusion characterization, illegal action discrimination and threshold value generation and event record output aiming at the aspect of complex shielding illumination change conditions in real time discrimination of the personnel illegal action of the chemical industry park. According to the embodiment of the invention, the chemical industry park personnel illegal action monitoring system based on pattern recognition comprises the following modules: The video acquisition module acquires continuous video streams of a chemical industry park operation