CN-121330738-B - Motion fatigue state non-contact detection method based on artificial intelligence

CN121330738BCN 121330738 BCN121330738 BCN 121330738BCN-121330738-B

Abstract

The invention provides a non-contact detection method for a sports fatigue state based on artificial intelligence, and relates to the field of non-contact sports fatigue detection. The method comprises the steps of analyzing facial visible light videos of users to realize non-contact assessment of motion fatigue states, innovatively providing a knowledge-guided nonlinear factorization alignment and projection framework, firstly obtaining semantic concept normal vectors representing semantic directions of core medical indexes based on motion fatigue knowledge maps, then carrying out nonlinear alignment conversion on global visual features by using an alignment network corresponding to each core medical index one by one, carrying out dot product projection on each alignment result and the corresponding semantic concept normal vector, splicing each projection result, effectively mapping high-dimensional and semantic entangled depth visual features into low-dimensional and structured physiological domain knowledge concepts, and providing a dynamic factorization hidden Markov model capable of deeply understanding motion fatigue accumulation effects and capturing dynamic evolution rules of the dynamic Markov model.

Inventors

DING SHUAI
CUI HAITAO
SONG CHENG
LIU ZIHAO
HU JUNZHE

Assignees

合肥工业大学

Dates

Publication Date: 20260512
Application Date: 20250930

Claims (9)

1. The non-contact detection method for the sports fatigue state based on the artificial intelligence is characterized by comprising the following steps of: Collecting face visible light videos of a user in a resting state, slicing, embedding each video slice into a corresponding visual representation, and acquiring a single global visual representation in an aggregation mode; Based on the constructed exercise fatigue knowledge graph, semantic concept normal vectors are obtained to represent semantic directions of different core medical indexes for evaluating exercise fatigue, wherein the core medical indexes at least comprise resting heart rate increase, heart rate variability decrease, anxiety emotion, depression emotion, irritability emotion and drowsiness; Based on the current global visual representation, nonlinear alignment conversion is carried out by using an alignment network corresponding to each core medical index one by one, dot product projection is carried out on each alignment result and a corresponding semantic concept normal vector, and each projection result is spliced to obtain a current interpretable medical index sequence, wherein a nonlinear multi-layer perceptron is selected as the alignment network; Based on the current sequence of interpretable medical indicators, the user's exercise fatigue state is assessed using a dynamic factorized hidden Markov model.
2. The method for non-contact detection of a sports fatigue state according to claim 1, wherein the construction process of the sports fatigue knowledge graph comprises the following steps: counting the reasons and symptoms of sports fatigue, wherein the symptoms at least comprise the core medical index; Constructing a first triplet (head node-relation-tail node) by taking any one of the reasons as a head node and the sports fatigue as a tail node or taking the sports fatigue as a head node and any one of the symptoms as a tail node, and acquiring an initial sports fatigue knowledge graph; And based on the initial exercise fatigue knowledge graph, supplementing and constructing a second triplet (head node-relation-tail node) for representing the association relation between every two of the core medical indexes so as to acquire the final exercise fatigue knowledge graph.
3. The method for non-contact detection of a state of athletic fatigue of claim 1, wherein the dynamically factored hidden Markov model comprises a historical encoder, a transition network, an emission network, and a hidden Markov model, wherein the evaluating the athletic fatigue of the user using the dynamically factored hidden Markov model based on the current sequence of interpretable medical metrics comprises: generating a history context vector using the history encoder based on history information of a current sequence of interpretable medical indicators, wherein the history information refers to a depth sequence consisting of all historic interpretable medical indicator sequences that cut off a current time step; Mapping the historical context vector into a dynamic transition probability matrix by utilizing the transition network based on the historical context vector; coding hidden variable fatigue states based on the historical context vectors, and acquiring the mean value and standard deviation of one-dimensional Gaussian distribution of each core medical index under different hidden variable fatigue states by utilizing a transmitting network corresponding to each core medical index one by one, wherein the hidden variable fatigue states correspond to preset sports fatigue types one by one; Based on the dynamic transition probability matrix, combining the mean value and standard deviation of one-dimensional Gaussian distribution of each core medical index in different hidden variable fatigue states, and calculating joint emission probability of observing a current interpretable medical index sequence in different hidden variable fatigue states by using the hidden Markov model; And determining the maximum joint emission probability and the corresponding hidden variable fatigue state thereof, and taking the sports fatigue type corresponding to the hidden variable fatigue state as a sports fatigue state evaluation result of the user.
4. A method for non-contact detection of a state of exercise fatigue according to claim 3, wherein a sequence-level conditional random field loss is used as a core training target, the conditional random field loss being expressed as: wherein the subscript Represents a conditional random field; As a logarithmic function; Representing a set of all possible hidden variable fatigue state sequences, Representing any one possible hidden variable fatigue state sequence; Is an exponential function; Representing a path calculated non-normalized logarithmic score; representing a given sequence of interpretable medical indicators, Representing the sequence length; Representing a real fatigue state; representing any one path; representing a dependency on an initial historical context vector Sequence-initiated hidden variable fatigue state Log probability of (a); representation relies on historical context vectors From the state of hidden variable fatigue To the point of Log transition probabilities of (a); representation relies on historical context vectors In the state of hidden variable fatigue Down-emission of an observed sequence of interpretable medical indicators Is a logarithmic transmission probability of (c).
5. The method for non-contact detection of sports fatigue according to any of claims 1-4, wherein TimeSformer is selected to embed the video slice into a visual representation.
6. The method for non-contact detection of athletic fatigue of claim 3 or 4, Selecting a multi-layer cyclic neural network as the history encoder; and/or selecting a nonlinear multi-layer perceptron as the transfer network; and/or selecting a multi-layer perceptron as the transmitting network.
7. An artificial intelligence based sports fatigue state non-contact detection system for performing the sports fatigue state non-contact detection method according to claim 1, comprising: The data processing module is used for collecting facial visible light videos in a user resting state and slicing the videos, and embedding and processing each video slice into a corresponding visual representation so as to obtain a single global visual representation in an aggregation mode; the pattern embedding module is used for acquiring semantic concept normal vectors based on the constructed exercise fatigue knowledge patterns so as to represent semantic directions of different core medical indexes for evaluating exercise fatigue, wherein the core medical indexes at least comprise resting heart rate increase, heart rate variability reduction, anxiety emotion, depression emotion, irritability emotion and drowsiness; The feature extraction module is used for carrying out nonlinear alignment conversion by using an alignment network corresponding to each core medical index one by one based on the current global visual representation, carrying out dot product projection on each alignment result and a corresponding semantic concept normal vector, and splicing each projection result to obtain a current interpretable medical index sequence; And the state evaluation module is used for evaluating the motion fatigue state of the user by utilizing a dynamic factorization hidden Markov model based on the current interpretable medical index sequence.
8. A storage medium storing a computer program for non-contact detection of a sports fatigue state based on artificial intelligence, wherein the computer program causes a computer to execute the sports fatigue state non-contact detection method according to any one of claims 1 to 6.
9. An electronic device, comprising: The apparatus of any one of claims 1-6, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs comprising instructions for performing the athletic fatigue status non-contact detection method of any one of claims 1-6.

Description

Motion fatigue state non-contact detection method based on artificial intelligence Technical Field The invention relates to the field of non-contact type sports fatigue detection, in particular to a sports fatigue state non-contact type detection method based on artificial intelligence. Background Sports fatigue is one of physiological parameters, is an important medical index for reflecting the functional state of a human body, and how to perform non-contact sports fatigue detection becomes a current hot problem. With the wide application of deep learning in the field of visual analysis, related technology has achieved the extraction of high-dimensional, rich biological visual features from facial visible light videos for detecting sports fatigue. In addition, in order to further improve the data efficiency, generalization capability and interpretability of the deep learning model, the related technology also fully utilizes various prior knowledge fields existing in the target field so as to make up for the defects of the traditional deep learning. For example, paper (Feng Wei,Jucheng Yang,Yuan Wang,Liang Lin,Haibin Zhang,Prior knowledge-guided multi-information graph convolutional network for driver drowsiness detection,Expert Systems with Applications,Volume 275,2025,127028,ISSN 0957-4174.) proposes a priori knowledge guided multi-information graph convolution network (MIGCN) to address the drowsiness detection problem of the driver. Compared to the CNN model-based driver drowsiness detection method, the structure MIGCN can effectively learn spatial facial features, enhancing feature representation. However, the above-mentioned technical solution does not effectively connect the high-dimensional deep visual features of the black box with the low-dimensional structured domain knowledge concepts with explicit physiological meanings, which greatly limits the application potential of the deep learning model in sports fatigue assessment. Disclosure of Invention (One) solving the technical problems Aiming at the defects of the prior art, the invention provides a non-contact detection method for the motion fatigue state based on artificial intelligence, which solves the technical problem of how to connect the high-dimension depth visual characteristics of a black box with the low-dimension structured domain knowledge concepts with definite physiological significance. (II) technical scheme In order to achieve the above purpose, the invention is realized by the following technical scheme: A non-contact detection method for motion fatigue state based on artificial intelligence comprises the following steps: Collecting face visible light videos of a user in a resting state, slicing, embedding each video slice into a corresponding visual representation, and acquiring a single global visual representation in an aggregation mode; Based on the constructed exercise fatigue knowledge graph, obtaining a plurality of semantic concept normal vectors to represent semantic directions of different core medical indexes for evaluating exercise fatigue, wherein the core medical indexes at least comprise resting heart rate rise, heart rate variability reduction, anxiety emotion, depression emotion, irritability emotion and drowsiness; based on the current global visual representation, nonlinear alignment conversion is carried out by using an alignment network corresponding to each core medical index one by one; performing dot product projection on each alignment result and the corresponding semantic concept normal vector, and splicing each projection result to obtain a current interpretable medical index sequence; Based on the current sequence of interpretable medical indicators, the user's exercise fatigue state is assessed using a dynamic factorized hidden Markov model. Preferably, the construction process of the exercise fatigue knowledge graph comprises the following steps: counting the reasons and symptoms of sports fatigue, wherein the symptoms at least comprise the core medical index; Constructing a first triplet (head node-relation-tail node) by taking any one of the reasons as a head node and the sports fatigue as a tail node or taking the sports fatigue as a head node and any one of the symptoms as a tail node, and acquiring an initial sports fatigue knowledge graph; And based on the initial exercise fatigue knowledge graph, supplementing and constructing a second triplet (head node-relation-tail node) for representing the association relation between every two of the core medical indexes so as to acquire the final exercise fatigue knowledge graph. . Preferably, each core medical index in the exercise fatigue knowledge graph is learned into a corresponding semantic concept normal vector by using a knowledge graph embedding method RotatE. Preferably, the dynamic factorization hidden Markov model comprises a history encoder, a transfer network, a transmitting network and a hidden Markov model, wherein the dynamic factoriza