CN-121982779-A - Deep learning-based water area personnel dangerous behavior identification method

CN121982779ACN 121982779 ACN121982779 ACN 121982779ACN-121982779-A

Abstract

The invention discloses a water area personnel dangerous behavior identification method based on deep learning, which relates to the technical field of video monitoring and comprises the steps of inhibiting image interference through Gaussian filtering and extracting human body posture key points, calculating a space-time motion vector field based on an optical flow method to generate a limb motion evolution sequence, fusing posture and motion characteristics when the sequence shows irregular fluctuation, judging whether a track deviates from a normal mode through a convolutional neural network to determine dangerous signals, extracting head water entering frequency and limb coordination degree indexes, adopting a support vector machine to classify to obtain abnormal behavior probability distribution, and triggering an alarm and outputting behavior type identification in real time according to the probability distribution crossing threshold. The method and the device effectively improve the accuracy and the instantaneity of dangerous behavior identification.

Inventors

ZHOU JIAHAO
LIU YU
XU HAORAN
Wei Yushu
WANG YUHAN

Assignees

济宁学院

Dates

Publication Date: 20260505
Application Date: 20260205

Claims (10)

1. The method for identifying dangerous behaviors of water area personnel based on deep learning is characterized by comprising the following steps of: Acquiring water area scene image data and extracting human body posture information; Based on the human body posture information, calculating space-time motion characteristics of a multi-frame sequence, and generating a limb action evolution sequence; Responding to the limb action evolution sequence to present irregular fluctuation, fusing human body posture information and space-time motion characteristics, and judging whether the motion trail deviates from a normal mode or not through a convolutional neural network so as to determine a potential dangerous signal; Extracting head water entering frequency and limb coordination degree indexes in the potential dangerous signals, and classifying the indexes by adopting a support vector machine to obtain abnormal behavior probability distribution; And carrying out threshold monitoring based on the abnormal behavior probability distribution, triggering an alarm when the abnormal behavior probability distribution crosses a preset threshold, and outputting a dangerous behavior type identifier.
2. The method of claim 1, wherein the process of acquiring water scene image data and extracting human body posture information comprises: Performing Gaussian filtering on an original image acquired by a video sequence to inhibit water surface reflection and wave interference; Performing edge detection and morphological refinement treatment on the filtered image to obtain a human body contour skeleton; And identifying the end points and bifurcation points of the human body contour skeleton to determine the positions of the key points of the preliminary gesture.
3. The method of claim 2, wherein computing the spatiotemporal motion characteristics of the sequence of frames based on the body pose information comprises: Calculating pixel displacement between adjacent frames by using the preliminary gesture key points as centers and utilizing an optical flow method to generate a space-time motion vector field; Abnormal vector elimination and interpolation processing are carried out on the space-time motion vector field, and a corrected vector field is obtained; and carrying out integral operation on the corrected vector field, constructing a continuous track mode reflecting the time dimension change of the key points, and generating a limb action evolution sequence based on the continuous track mode.
4. A method according to claim 3, wherein the process of fusing human body posture information with spatiotemporal motion characteristics and determining by convolutional neural network comprises: Fusing the gesture key point sequence corresponding to the irregular fluctuation period with the space-time motion vector to construct a multidimensional fusion characteristic tensor; inputting the multidimensional fusion feature tensor into a convolutional neural network to extract a deep feature map and generate an actual motion track; And calculating the distance between the actual movement track and the predefined normal swimming track manifold, and determining a potential danger signal according to whether the distance exceeds a threshold value.
5. The method of claim 4, wherein the process of extracting the index of head entry frequency and limb coordination comprises: based on the time sequence data stream of skeleton key points output by the gesture estimation network, analyzing the vertical displacement fluctuation of head key points to count the head water inlet frequency; calculating the change rate of the angles of the joints of the limbs according to the time sequence data stream so as to evaluate limb coordination; and vectorizing and splicing the head water inlet frequency index and the limb coordination index to form a multidimensional fusion feature vector.
6. The method of claim 5, wherein the process of threshold monitoring and triggering an alarm based on the behavioral anomaly probability distribution comprises: constructing an abnormal probability value obtained through real-time calculation as an abnormal probability distribution matrix; Extracting statistical characteristic values of the abnormal probability distribution matrix, and sending a trigger instruction when the statistical characteristic values exceed a preset monitoring threshold value; And the alarm mechanism matches the distribution characteristics of the current behaviors with the risk behavior characteristic library according to the triggering instruction, and determines and outputs a specific dangerous behavior type identifier.
7. A deep learning-based water personnel dangerous behavior identification system for implementing the method of any of claims 1-6, the system comprising: The data acquisition and gesture extraction module is used for acquiring water area scene image data and extracting human body gesture information; the motion analysis module is used for calculating the space-time motion characteristics of the multi-frame sequence based on the human body posture information and generating a limb motion evolution sequence; The risk identification module is used for responding to the limb action evolution sequence to present irregular fluctuation, fusing human body posture information and space-time motion characteristics, judging whether the motion trail deviates from a normal mode or not through a convolutional neural network, and determining potential dangerous signals; the behavior classification module is used for extracting head water entering frequency and limb coordination degree indexes in the potential dangerous signals, and classifying the indexes by adopting a support vector machine to obtain behavior anomaly probability distribution; And the alarm triggering module is used for carrying out threshold monitoring based on the abnormal behavior probability distribution, triggering an alarm when the threshold value crosses a preset threshold value and outputting a dangerous behavior type identifier.
8. A computer terminal device, comprising: One or more processors; A memory coupled to the processor for storing one or more programs; When executed by the one or more processors, causes the one or more processors to implement the steps of the method of any of claims 1-6.
9. A computer readable storage medium, on which a computer program is stored, characterized in that the computer program, when being executed by a processor, implements the steps of the method according to any of claims 1-6.
10. A computer program product comprising a computer program, characterized in that the computer program, when being executed by a processor, implements the steps of the method according to any of claims 1-6.

Description

Deep learning-based water area personnel dangerous behavior identification method Technical Field The invention belongs to the technical field of video monitoring, and particularly relates to a water area personnel dangerous behavior identification method based on deep learning. Background The water area safety monitoring is an important field for guaranteeing public life safety, particularly in open water areas such as swimming pools, lakes, rivers, beaches and the like, timely identification and early warning of dangerous behaviors such as personnel falling into water or drowning occur, and the rescue efficiency and the survival chance are directly related. Currently, automatic recognition technology based on computer vision is commonly adopted in the field, and the core of the automatic recognition technology is mostly dependent on a traditional image processing method or a machine learning model based on single-mode characteristics. For example, the moving object is detected by background difference or interframe difference method, and then the behavior state is judged by combining the human contour shape or simple track tracking. However, when deployed in a practical complex natural water surface environment, such methods pose serious challenges in terms of reliability and accuracy. The primary difficulties arise from strong specular reflection from the water surface and persistent non-rigid wave disturbances. The strong reflection of light can produce a highlight region in the image, severely overexposure and even complete masking of the contours of the human body parts, while the constantly changing wave texture can introduce a large amount of random noise and dynamic artifacts in the background and foreground. These environmental disturbances can significantly distort and obscure the visual information of the human body, resulting in the human body contour extraction algorithm based on threshold values or gradients becoming extremely unstable and difficult to obtain a clear and consistent human body appearance. Distortion and breakage of the contours further cause serious deviations or loss of subsequent human body pose estimates or key point positioning. The fundamental defect is that the prior art scheme often analyzes static attitude characteristics or low-order motion characteristics in an isolated manner, and fails to effectively integrate complete behavior information of a human body in space-time dimension. On one hand, the disturbed and unstable static attitude information (such as joint coordinates) has low reliability, and on the other hand, the motion characteristics obtained by simply calculating the displacement of the mass center of the target or a simple optical flow field cannot accurately describe the fine action mode of the limb, such as disordered struggling or regular rowing of the arm. The cleavage of the gesture information and the deep space-time motion characteristics makes it difficult for the system to capture the essential characteristics of dangerous behaviors (such as drowning), namely the coupling of abnormal limb gestures and irregular space-time motion modes. In an actual scene, the system is often caused to misjudge the profile shake of a normal swimmer caused by stormy waves as struggling, or the behavior of a person with a slight drowning sign (such as frequent sinking of the head into the water and uncoordinated actions) is caused to be missed due to incomplete feature extraction, so that an error alarm is caused or a critical rescue opportunity is delayed. How to robustly extract and effectively integrate human body posture and fine space-time motion characteristics under the strong interference environment of the water surface so as to realize high-precision and high-reliability real-time identification of dangerous behaviors such as drowning and the like, and the method becomes a core technical problem to be solved urgently in the level of the safety monitoring intelligence of the lifting water area. Disclosure of Invention In order to solve the technical problems, the invention provides a water area personnel dangerous behavior identification method based on deep learning, which aims to solve the problems existing in the prior art. In order to achieve the above object, the present invention provides a method for identifying dangerous behavior of a water area person based on deep learning, comprising the following steps: Acquiring water area scene image data and extracting human body posture information; Based on the human body posture information, calculating space-time motion characteristics of a multi-frame sequence, and generating a limb action evolution sequence; Responding to the limb action evolution sequence to present irregular fluctuation, fusing human body posture information and space-time motion characteristics, and judging whether the motion trail deviates from a normal mode or not through a convolutional neural network so as to determine a potent