Police martial arts fighting action evaluation method and system based on peak shifting time pulse

Abstract

The invention discloses a police martial arts fighting action evaluation method and system based on peak shifting time pulses, belongs to the technical field of action behavior analysis, and solves the problem that tiny dynamic changes of joints cannot be captured in time when an existing method processes a complex motion sequence. The method comprises the steps of extracting key skeleton points from a target action video, converting serialized motion data into a discrete pulse signal set through a pulse function, and processing the pulse signal set based on a peak shifting time function. The multi-dimensional time sequence evaluation model carries out police martial arts fighting action recognition and training evaluation on the target action video based on the pulse feature set; according to the method, the time sequence relation of the skeleton points is analyzed by introducing the peak shifting time function, the time sequence and integrity of feature extraction are ensured based on the feature adaptive extraction algorithm of the flexible trigger mechanism of the water tank, and the accuracy and robustness of motion behavior evaluation are improved, so that a scientific and effective evaluation tool is provided for motion training and evaluation.

Description

Technical Field

[0001] The present invention belongs to the technical field of sports behavior analysis, and specifically relates to a method and system for evaluating police martial arts fighting actions based on staggered time pulses. Background Art

[0002] In martial arts training, it is crucial to evaluate the accuracy of training fighting actions. Currently, traditional evaluation methods mainly rely on two ways: coach observation and video playback. The coach observation method is an intuitive evaluation method, in which the coach makes a subjective judgment on the actions based on experience. Although this method is simple and direct, due to the lack of a quantitative standard, the accuracy and consistency of its evaluation results are difficult to guarantee, and it is subject to the personal ability and attention of the coach. In contrast, the video playback method records the training footage and analyzes the action details frame by frame. This method can capture more detailed information, but it also has problems such as relying on manual analysis, low efficiency, and inability to provide real-time feedback, making it difficult to meet the needs of rapid evaluation.

[0003] With the development of technology, action recognition methods based on sensors and optical capture have gradually been applied to training evaluation. These methods use modern technical means to quantitatively analyze action data, significantly improving the evaluation accuracy and scientificity. Among them, the sensor-based action capture method collects human action data in real time by wearing devices such as accelerometers and gyroscopes, and analyzes the action parameters in combination with biomechanical models. This method can capture action details with high accuracy, but the sensor devices are expensive and need to be worn, which causes certain interference to the naturalness of training. At the same time, it has high requirements for the installation and debugging of the devices and the technical level of the operators, limiting its practical application in high-intensity training scenarios. Another type is the optical-based action capture method, which uses reflective markers and a multi-camera system to capture the human movement trajectory, generate a skeletal model and conduct a detailed analysis. This method is widely used in the scientific research field and can accurately capture the action trajectory and posture. However, its high equipment cost, complex environmental requirements, and dependence on specific optical conditions make it difficult to be popularized in actual training scenarios. In addition, the optical capture system is easily affected by external interference in dynamic scenarios, such as light changes or actions exceeding the camera capture range, further limiting its applicability.

[0004] In recent years, video analysis technology based on skeletal point data has gradually become the mainstream in the field of action recognition. Such methods extract the human skeletal key point data in the video through algorithms and analyze the actions by combining rule inference or key frame extraction techniques. Although the automation level has been significantly improved compared with traditional means, there are still many technical bottlenecks. First of all, existing methods usually rely on preset empirical thresholds or simple rule algorithms to identify key frames in actions. This method may perform well in static or simple scenarios, but in complex motion environments or for action sequences with large individual differences, the robustness and adaptability are insufficient, resulting in limited recognition accuracy and stability for high-dynamic change actions. Secondly, due to the high-dimensional characteristics of human skeletal point data, there is a trade-off between data dimension and computational efficiency during the feature extraction process. To reduce complexity, researchers often simplify the processing by dimensionality reduction or selecting some key points. However, this method is prone to losing important motion information, affecting the integrity of feature extraction and the overall recognition effect. In addition, police martial arts fighting actions often involve complex dynamic behaviors, and the time staggering characteristics between their skeletal points are particularly significant. For example, different limbs have obvious timing differences when performing coherent actions, while existing action recognition methods usually rely on synchronous processing or simple timing modeling and are difficult to accurately represent the dynamic time relationship between action components. This limitation directly leads to insufficient feature mining of complex fighting actions and cannot meet the actual needs of high-precision action evaluation.

[0005] Chinese Patent CN118718339A discloses a motion assistance system and a motion assistance method. The system evaluates the video information related to the user's motion through a motion evaluation unit, calculates a motion score, and outputs a feature map to represent the contribution degree of joint parts. This method mainly relies on video analysis for motion evaluation. However, when dealing with complex motion sequences, it may face certain limitations and cannot capture the subtle dynamic changes of joints in a timely manner.

[0006] Chinese Patent CN118737376A discloses an information processing device, an information processing method, a program recording medium, and a model under learning, providing an information processing device, an information processing method, a program recording medium, and a model under learning, which can output age-appropriate guidance reports for each subject and can assist the subject's movement. The information processing device has: a subject information acquisition unit that acquires subject information including age information indicating the subject; a motion evaluation information acquisition unit that acquires motion evaluation information related to the motion evaluation of the subject; and a storage unit that stores a learned model. However, when dealing with complex motion sequences, the above method cannot timely capture the subtle dynamic changes of joints, nor can it provide a finer-grained motion analysis. At the same time, the existing methods are inefficient, lack robustness, adaptability, accuracy, and stability. Moreover, the high-dimensional characteristics of human skeleton point data make it difficult for the existing technology to balance data dimension and computational efficiency when extracting features. Complex action behaviors also show an obvious time staggering relationship between the skeleton points of different parts. Existing action recognition methods often ignore this feature. In view of the above problems, we propose a method and system for evaluating police martial arts fighting actions based on staggered time pulses. Summary of the Invention

[0007] The purpose of the present invention is to provide a method and system for evaluating police martial arts fighting actions based on staggered time pulses in view of the deficiencies of the existing technology, solving the problems that the existing methods cannot timely capture the subtle dynamic changes of joints and cannot provide a finer-grained motion analysis when dealing with complex motion sequences. At the same time, the existing methods are inefficient, lack robustness, adaptability, accuracy, and stability. Moreover, the high-dimensional characteristics of human skeleton point data make it difficult for the existing technology to balance data dimension and computational efficiency when extracting features. Complex action behaviors also show an obvious time staggering relationship between the skeleton points of different parts. Existing action recognition methods often ignore this feature. The present invention improves accuracy through data preprocessing, enhances reliability through filtering, generates pulse signals to emphasize temporal features, analyzes the temporal relationship of skeleton points through a staggered time function, and ensures the temporality and integrity of feature extraction through a feature adaptive extraction algorithm with a flexible trigger mechanism for the water tank.

[0008] The present invention is implemented as follows. The method for evaluating police martial arts fighting actions based on staggered time pulses includes:

[0009] Obtain a target action video, extract key skeleton points from the target action video, number the key skeleton points, and cluster the skeleton points with corresponding codes based on regional anatomical features;

[0010] Obtain a serialized clustering center point sequence, convert the serialized motion data into a discrete set of pulse signals through a pulse function, and perform serialization processing on the set of pulse signals;

[0011] Load a set of pulse signal, adjust the weights of the pulse signal set based on the peak staggering time function, and put the pulse signal set with adjusted weights into a flexible triggering mechanism for processing to obtain a pulse feature set;

[0012] Construct and train a multi-dimensional time series evaluation model, use the pulse feature set as the input, execute the multi-dimensional time series evaluation model, and the multi-dimensional time series evaluation model performs police martial arts fighting action recognition and training evaluation on the target action video based on the pulse feature set, and outputs the action recognition and training evaluation results of the target action video.

[0013] The method for extracting key skeleton points from the target action video includes:

[0014] Load at least one set of target action videos, and convert the target action videos frame by frame into at least one set of action images;

[0015] Use the human pose estimation algorithm BlazePose to detect key skeleton points for each frame of the action image, detect the key skeleton points of the human body, and number the key skeleton points. Each detected skeleton point is assigned a fixed number i;

[0016] Among them, when the human pose estimation algorithm BlazePose detects key skeleton points for each frame of the action image, it analyzes the human body contour and joint points in the image, estimates the coordinates of each skeleton point in three-dimensional space, and assumes that at time t, the coordinates of the skeleton point numbered i are:

[0017] p i (t) = (x i (t), y i (t), z i (t))

[0018] Among them, (x i (t), y i (t), z i (t)) represents the coordinates of the skeleton point i in three-dimensional space;

[0019] Obtain the key skeleton points of the detected human body, and perform filtering and noise reduction processing on the key skeleton points of the detected human body based on a low-pass filter;

[0020] The filtered coordinates are calculated by the following formula:

[0021]

[0022] When clustering the skeleton points with corresponding encodings based on regional anatomical features, the clustering result is expressed as:

[0023]

[0024] Among them, each cluster C k represents a set of related skeletal points, reflecting the common characteristics existing among them during movement.

[0025] The method of converting serialized motion data into a discrete set of pulse signals through a pulse function includes:

[0026] Loading the sequence of clustering center points c k (t), calculating the velocity and performing pulse processing on the sequence of clustering center points to extract the temporal characteristics of the action;

[0027] Let v k (t) be the velocity of the clustering center point c k (t) at time t, defined as:

[0028]

[0029] where Δt is the time interval between adjacent sampling times;

[0030] Presetting a velocity threshold function θ k for generating pulse signals;

[0031] When the velocity of the clustering center point exceeds the preset velocity threshold function θ k at a certain moment, it is considered to be in an active state at that moment, and a corresponding pulse signal s k (t) is generated:

[0032]

[0033] where s k (t) is the pulse signal of the k-th cluster at time t;

[0034] Integrating at least one set of pulse signals s k (t) to obtain a discrete set of pulse signals;

[0035] Loading the set of pulse signals, serializing the pulse signals of each group of clustering center points to obtain pulse signals in the form of a time series:

[0036] S k ={s k (1),s k (2),…,s k (T)}

[0037] where T is the time length of the entire sequence;

[0038] Defining the weighted fusion pulse signal S f (t):

[0039]

[0040] Among them, w k represents the weight of the k-th cluster, which is used to balance the contributions of different clusters to highlight the clusters that dominate the action features. In this process, by weighted fusion of the pulse signals of different clustering clusters, a unified pulse sequence S f (t) representing the overall action is obtained.

[0041] The method of adjusting the weights of the pulse signal set based on the stagger time function and processing the adjusted pulse signal set in the flexible trigger mechanism includes:

[0042] Loading the pulse signal set, based on the stagger time function, identifying the movement order between multiple skeleton points in the pulse signal set according to the movement time difference between skeleton points, capturing the dynamic features of the target action, and identifying the occurrence of coherent actions through the analysis of the time series features between multiple skeleton point categories;

[0043] Among them, the stagger time function τ ij (t) is defined as:

[0044] τ ij (t) = exp(-α|(t j -t i ) - Δt ij |)

[0045] Among them, t i and t j are the timestamps of clusters C i and C j at the t-th moment respectively, α is a parameter controlling the attenuation of the function, and it exponentially decays through the deviation of the time series of the clustering center point. When the time difference is closer to the preset delay Δt ij , the function value is larger, indicating that the action timing meets the expectation; on the contrary, if the deviation is large, the function value tends to zero, indicating that there is an abnormal timing in the action;

[0046] Taking the output of the function τ ij (t) as the dynamic weight and applying it to the pulse signal S f (t):

[0047]

[0048] Among them is the adjusted pulse signal, and the time difference feature is dynamically enhanced by the stagger time weight τ ij (t).

[0049] Loading the pulse signal set processed by the stagger time function Adaptive extraction of the characteristics of a pulse signal set is combined with a flexible trigger mechanism based on a water tank, and a pulse feature set is output.

[0050] Load the pulse signal set processed by the stagger time function, combine the adaptive extraction of the characteristics of the pulse signal set with a flexible trigger mechanism based on a water tank, and output a pulse feature set.

[0051] The flexible trigger mechanism based on the water tank uses a nested water tank model to process the pulse signal set;

[0052] When the nested water tank model processes the pulse signal set, a small water tank state function Q k (t) is defined. The water tank state function Q k (t) represents the cumulative value of the small water tank at time t. The cumulative rule of the water tank state function Q k (t) is as follows:

[0053]

[0054] Among them, represents the input of the pulse signal from the adjusted clustering cluster C k after being adjusted by the stagger time function. λ t is the linear time decay coefficient of the water tank, which is used to control the automatic emptying speed of the water tank;

[0055] To ensure that the data in the small water tank represents a complete action, a cumulative threshold η k is defined. When Q k (t) ≥ η k , it is considered that the action is completed and the feature is output:

[0056]

[0057] After the feature is output, the small water tank will be reset:

[0058] Q k (t) = 0

[0059] If there is not enough pulse signal input within a period of time, the cumulative value in the small water tank will gradually decrease with time.

[0060] The nesting mechanism of the nested water tank model is expressed as:

[0061] Multiple small water tanks are nested in a large water tank. The cumulative rule of the state function Q D (t) of the large water tank is:

[0062]

[0063] Among them, I(Q k (t) ≥ η k) is the output marking function of the small water tank. When the small water tank outputs a complete action, the value is 1; otherwise, it is 0; γ t is the linear attenuation coefficient of the large water tank, and the cumulative threshold of the large water tank is η D , when Q D (t) ≥ η D , it is considered that the whole set of actions is completed, the characteristics of the whole set of actions are output, and the large water tank is reset:

[0064] Q D (t) = 0

[0065] When the cumulative value of the water tank reaches the threshold, the output feature can be expressed as the overall feature F(t) of the serialized pulses of the skeleton points:

[0066] F(t) = {S output (t) | Q D (t) ≥ η D}

[0067] Among them, F(t) represents the feature output at time t, and S output (t) is the serialized pulse signal processed by the time staggering function and the water tank mechanism.

[0068] The multi-dimensional time series evaluation model is a spiking neural network, which consists of an input layer, multiple hidden layers, a pooling layer and an output layer. The input layer of the network receives the overall feature F(t) of the spiking feature set;

[0069] Among them, the input layer receives the overall feature F(t) as input, and is used to capture the motion pattern of the skeleton points in the time series;

[0070] Each hidden layer corresponds to a specific skeleton point category. The hidden layer processes information through the spiking neuron model. Using the spiking neuron model, the membrane potential u i (t) accumulates according to the input features, and the update rule is:

[0071] u i (t) = u i (t - 1) + w i F(t) - τu i (t - 1)

[0072] Among them, w i is the connection weight, and τ is the time constant that controls the decay of the membrane potential.

[0073] The neuron spiking mechanism of the multi-dimensional time series evaluation model is:

[0074] When the membrane potential u i (t) exceeds the threshold θ iWhen a neuron outputs a pulse signal, the output signal is defined as:

[0075]

[0076] Here, the output signal o(t) represents the pulse firing state of the neuron at time t, which is used to reflect the occurrence of the movement behavior;

[0077] Between the hidden layer and the output layer, there is a pooling layer. The pooling layer is used to integrate the output features from multiple hidden neurons and reduce the dimension of the data. The pooling layer uses max pooling to select the maximum value from the outputs of multiple spiking neurons to ensure that the most significant features are focused on in action recognition;

[0078] The pooling operation of the pooling layer is defined as:

[0079] P(t) = max(o1(t), o2(t),..., o n (t))

[0080] where P(t) is the output of the pooling layer, and o n (t) is the pulse output of the nth neuron in the hidden layer;

[0081] The output layer is responsible for converting the pulse signal of the hidden layer into the final action recognition result. The calculation method of the output feature is as follows:

[0082]

[0083] where w oi is the connection weight between the output layer neuron and the hidden layer neuron, and N is the number of hidden layer neurons.

[0084] When training the multi-dimensional time series evaluation model, the cross-entropy loss function is used to optimize the performance of the spiking neural network. The loss function of the multi-dimensional time series evaluation model is defined as:

[0085]

[0086] The connection weight w oi in the network is adjusted through the backpropagation algorithm to minimize the loss.

[0087] On the other hand, the present invention also provides an action evaluation system based on staggered time pulses. The action evaluation system based on staggered time pulses includes:

[0088] A data processing module, which is used to obtain the target action video, extract the key skeleton points from the target action video, number the key skeleton points, and cluster the skeleton points with corresponding codes based on the regional anatomical features;

[0089] A pulse processing module that converts serialized motion data into a discrete set of pulse signals through a pulse function and serially processes the set of pulse signals;

[0090] A feature extraction module for loading the set of pulse signals, adjusting the weights of the set of pulse signals based on a staggered time function, and processing the set of pulse signals with adjusted weights in a flexible trigger mechanism to obtain a set of pulse features;

[0091] An action evaluation module for constructing and training a multi-dimensional time series evaluation model, using the set of pulse features as input, executing the multi-dimensional time series evaluation model, and the multi-dimensional time series evaluation model performing police martial arts fighting action recognition and training evaluation on the target action video based on the set of pulse features and outputting the action recognition and training evaluation results of the target action video.

[0092] Compared with the prior art, the embodiments of the present application mainly have the following beneficial effects:

[0093] In the present invention, by introducing a staggered time function to analyze the timing relationship of skeleton points, actions and continuous actions are judged according to the time series differences of different skeleton points and their categories. The feature adaptive extraction algorithm based on the flexible trigger mechanism of the water tank ensures the timing and integrity of feature extraction. By constructing and training a multi-dimensional time series evaluation model based on a pulse neural network, the cross-entropy loss function is used for training in the design of the multi-dimensional time series evaluation model, thereby optimizing the performance of the pulse neural network, improving the accuracy and robustness of motion behavior evaluation, and thus providing a scientific and effective evaluation tool for action training and evaluation, greatly improving the practicality and scientific nature of training evaluation. The lightweight design of the pulse neural network enables the system to still perform real-time data processing and action recognition at a high speed and accuracy under the condition of limited hardware resources. It also has a strong time evaluation ability and can monitor and analyze action dynamics in real time, especially in a training environment with high dynamic changes. It overcomes the problems that the existing methods cannot capture the subtle dynamic changes of joints in a timely manner when dealing with complex motion sequences and cannot provide more fine-grained motion analysis. At the same time, it also solves the problems of inefficiency, lack of robustness, self-adaptability, accuracy and stability of the existing methods. Moreover, the high-dimensional characteristics of human skeleton point data make it difficult for the existing technology to balance data dimension and computational efficiency in feature extraction.

[0094] In the present invention, the serialized motion data is converted into a discrete set of pulse signals through a pulse function. This pulse conversion method transforms the continuous motion of the time series into pulse data with specific time nodes, retaining the important timing features in the motion process and helping to capture the dynamic changes in complex motions.

[0095] In the present invention, a low-pass filter is adopted to eliminate high-frequency components and retain the low-frequency trend information in the bone point coordinates. The filtering can effectively remove the jitter caused by pose estimation errors and is also conducive to reflecting the basic form and change trend of human movements, which is of great significance for accurately describing human movements. BRIEF DESCRIPTION OF THE DRAWINGS

[0096] Figure 1 FIG. is a schematic implementation flowchart of a police martial arts combat action evaluation method based on staggered time pulses provided by the present invention.

[0097] Figure 2 FIG. shows a schematic implementation flowchart of a method for extracting key bone points from a target action video.

[0098] Figure 3 FIG. shows a schematic implementation flowchart of a method for converting serialized motion data into a discrete set of pulse signals through a pulse function.

[0099] Figure 4 FIG. shows a schematic implementation flowchart of a method for adjusting the weights of a set of pulse signals based on a staggered time function and processing the adjusted set of pulse signals in a flexible trigger mechanism.

[0100] Figure 5 FIG. shows a logical schematic diagram of a flexible trigger mechanism based on a water tank.

[0101] Figure 6 FIG. shows an identification flowchart of a multi-dimensional time series evaluation model.

[0102] Figure 7 FIG. is a schematic structural diagram of an action evaluation system based on staggered time pulses provided by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0103] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which this application belongs; the terms used in the specification of this application are only for the purpose of describing specific embodiments and are not intended to limit this application; the terms "including" and "having" and any variations thereof in the specification and claims of this application and the above drawings are intended to cover non-exclusive inclusion. The terms "first", "second", etc. in the specification and claims of this application or the above drawings are used to distinguish different objects and not to describe a specific order.

[0104] References to "embodiments" in this specification mean that a particular feature, structure, or characteristic described in connection with the embodiments can be included in at least one embodiment of the present application. The phrase appears in various places in the specification and does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment mutually exclusive with other embodiments. Those skilled in the art will understand explicitly and implicitly that the embodiments described herein can be combined with other embodiments.

[0105] When dealing with complex motion sequences, existing methods cannot capture the subtle dynamic changes of joints in a timely manner, nor can they provide a finer-grained motion analysis. To address the above problems, we propose a method and system for evaluating police martial arts fighting actions based on staggered time pulses. When the method is implemented, first, a target action video is obtained, key skeleton points are extracted from the target action video, and the key skeleton points are numbered. Then, the serialized motion data is converted into a discrete set of pulse signals through a pulse function, and the set of pulse signals is serialized. Based on the staggered time function, the weights of the set of pulse signals are adjusted. The set of pulse signals with adjusted weights is processed in a flexible trigger mechanism. Finally, a multi-dimensional time series evaluation model performs police martial arts fighting action recognition and training evaluation on the target action video based on the pulse feature set, and outputs the action recognition and training evaluation results of the target action video. In the embodiments of the present invention, by introducing a staggered time function to analyze the temporal relationship of skeleton points, actions and continuous actions are judged based on the time series differences of different skeleton points and their categories. The feature adaptive extraction algorithm based on the flexible trigger mechanism of the water tank ensures the temporality and integrity of feature extraction. By constructing and training a multi-dimensional time series evaluation model based on a pulse neural network, the cross-entropy loss function is used for training in the design of the multi-dimensional time series evaluation model, so as to optimize the performance of the pulse neural network, improve the accuracy and robustness of motion behavior evaluation, and thus provide a scientific and effective evaluation tool for action training and evaluation, greatly improving the practicality and scientific nature of training evaluation. The lightweight design of the pulse neural network enables the system to still perform real-time data processing and action recognition at a high speed and accuracy under the condition of limited hardware resources. It also has a strong temporal evaluation ability, can monitor and analyze action dynamics in real time, especially in a training environment with high dynamic changes. It overcomes the problems that existing methods cannot capture the subtle dynamic changes of joints in a timely manner and cannot provide a finer-grained motion analysis when dealing with complex motion sequences. It also solves the problems of inefficiency, lack of robustness, adaptability, accuracy, and stability of existing methods. Moreover, the high-dimensional characteristics of human skeleton point data make it difficult for existing technologies to balance data dimensions and computational efficiency when extracting features.

[0106] It should be noted that the method for evaluating police martial arts fighting actions based on staggered time pulses in the embodiments of the present invention can be used in, but not limited to, competitive sports, medical rehabilitation, industrial production, intelligent monitoring, fitness exercises, and police martial arts training. Exemplarily, during police martial arts training, action evaluation is used to analyze and improve the fighting skills and tactical actions of training personnel to improve fighting efficiency and survival ability. The present invention also supports the detailed analysis and evaluation of complex fighting action processes. This method improves the accuracy and efficiency of police martial arts fighting action analysis and can provide a solid technical support and application foundation for the fighting training evaluation of police officers.

[0107] The embodiments of the present invention provide a method for evaluating police martial arts fighting actions based on staggered time pulses. Figure 1 The schematic diagram of the implementation process of the method for evaluating police martial arts fighting actions based on staggered time pulses is shown. The method for evaluating police martial arts fighting actions based on staggered time pulses specifically includes:

[0108] S10. Obtain the target action video, extract the key skeleton points from the target action video, number the key skeleton points, and cluster the skeleton points with corresponding codes based on the regional anatomical features.

[0109] In this embodiment, the target action video is obtained through a camera, a video recorder, or a machine vision-based monitoring camera. The key skeleton points are extracted from the target action video, and each skeleton point is numbered. These skeleton points correspond to the main joints and movement nodes of the human body and can effectively capture the dynamic information during the movement process. The extracted skeleton point data will be used as the basis for analysis.

[0110] S20. Obtain the serialized clustering center point sequence, convert the serialized motion data into a discrete pulse signal set through a pulse function, and perform serialization processing on the pulse signal set.

[0111] In the embodiments of the present invention, the serialized motion data is converted into a discrete pulse signal set through a pulse function. This pulse conversion method converts the continuous motion of the time series into pulse data with specific time nodes, retains the important timing features during the movement process, and helps to capture the dynamic changes in complex movements.

[0112] S30. Load the pulse signal set, adjust the weight of the pulse signal set based on the staggered time function, and put the pulse signal set with adjusted weight into a flexible trigger mechanism for processing to obtain a pulse feature set.

[0113] S40. Construct and train a multi-dimensional time series evaluation model. Using the pulse feature set as the input, execute the multi-dimensional time series evaluation model. The multi-dimensional time series evaluation model performs police martial arts fighting action recognition and training evaluation on the target action video based on the pulse feature set, and outputs the action recognition and training evaluation results of the target action video.

[0114] In the embodiments of the present invention, by introducing a staggered time function to analyze the temporal relationship of skeleton points, to judge actions and continuous actions according to the time series differences of different skeleton points and their categories, the feature adaptive extraction algorithm based on the flexible trigger mechanism of the water tank ensures the temporality and integrity of feature extraction. By constructing and training a multi-dimensional time series evaluation model based on a pulse neural network, the cross-entropy loss function is used for training in the design of the multi-dimensional time series evaluation model, thereby optimizing the performance of the pulse neural network, improving the accuracy and robustness of motion behavior evaluation, and thus providing a scientific and effective evaluation tool for action training and evaluation, greatly enhancing the practicality and scientific nature of training evaluation. The lightweight design of the pulse neural network enables the system to still perform real-time data processing and action recognition at a high speed and accuracy under the condition of limited hardware resources. It also has a strong temporal evaluation ability, capable of real-time monitoring and analyzing action dynamics, especially in a training environment with high dynamic changes. It overcomes the problems that existing methods cannot capture the subtle dynamic changes of joints in a timely manner when dealing with complex motion sequences, nor can they provide finer-grained motion analysis. At the same time, it also solves the problems of inefficiency, lack of robustness, self-adaptability, accuracy, and stability of existing methods. Moreover, the high-dimensional characteristics of human skeleton point data make it difficult for existing technologies to balance data dimensions and computational efficiency when extracting features.

[0115] The embodiments of the present invention provide a method for extracting key skeleton points from a target action video. Figure 2 The figure shows a schematic implementation flow diagram of the method for extracting key skeleton points from a target action video. The method for extracting key skeleton points from a target action video specifically includes:

[0116] S101. Load at least one group of target action videos, and convert the target action videos frame by frame into at least one group of action images;

[0117] It should be noted that the input target action video data is a collected continuous video stream. In order to analyze the posture and movement of the human body in combat sports, it is necessary to convert the target action video frame by frame into at least one group of action images, and then extract the skeleton points in the video frames.

[0118] S102. Use the human pose estimation algorithm BlazePose to detect key skeleton points for each frame of action image, detect the key skeleton points of the human body, and number the key skeleton points. Each detected skeleton point is assigned a fixed number i;

[0119] Among them, when the human pose estimation algorithm BlazePose detects key skeleton points for each frame of action image, it analyzes the human body contour and joint points in the image, estimates the coordinates of each skeleton point in three-dimensional space. Assuming that at time t, the coordinates of the skeleton point numbered i are:

[0120] p i (t) = (x i (t), y i (t), z i (t))

[0121] Among them, (x i (t), y i (t), z i (t)) represents the coordinates of the skeleton point i in three-dimensional space;

[0122] S103. Obtain the key skeleton points of the detected human body, and perform filtering and noise reduction processing on the key skeleton points of the detected human body based on a low-pass filter;

[0123] The filtered coordinates are calculated by the following formula:

[0124]

[0125] When clustering the skeleton points with corresponding encodings based on regional anatomical features, the clustering result is expressed as:

[0126]

[0127] Among them, each cluster C k represents a group of related skeleton points, reflecting the common features they have during movement.

[0128] In this embodiment, in order to reduce noise and improve the reliability of data, it is necessary to filter the extracted skeleton point coordinates. In this application, a low-pass filter is used to eliminate high-frequency components and retain the low-frequency trend information in the skeleton point coordinates. Filtering can effectively remove the jitter caused by pose estimation errors, and is also conducive to reflecting the basic shape and change trend of human movements, which is of great significance for accurately describing human movements.

[0129] The embodiment of the present invention provides a method for converting serialized motion data into a discrete set of pulse signals through a pulse function, Figure 3 which shows a schematic implementation flowchart of the method for converting serialized motion data into a discrete set of pulse signals through a pulse function. The method for converting serialized motion data into a discrete set of pulse signals through a pulse function specifically includes:

[0130] S201, Load the sequence of clustering center points c k (t), calculate the velocity and pulse process the sequence of clustering center points, and extract the temporal features of the action;

[0131] Let v k (t) be the velocity of the clustering center point c k (t) at time t, defined as:

[0132]

[0133] where Δt is the time interval between adjacent sampling times;

[0134] S202, Preset a velocity threshold function θ k , and the velocity threshold function is used to generate a pulse signal;

[0135] S203, When the velocity of the middle clustering center point exceeds the preset velocity threshold function θ k , it is considered to be active at that moment, and a corresponding pulse signal s k (t) is generated:

[0136]

[0137] where s k (t) is the pulse signal of the k-th cluster at time t;

[0138] S204, Integrate at least one set of pulse signals s k (t) to obtain a discrete set of pulse signals;

[0139] S205, Load the set of pulse signals, serialize the pulse signals of each group of clustering center points, and obtain the pulse signals in the form of a time series:

[0140] S k ={s k (1), s k (2), …, s k (T)}

[0141] where T is the time length of the entire sequence;

[0142] Define the weighted fusion pulse signal S f (t):

[0143]

[0144] where, w kDenote the weight of the k-th cluster, which is used to balance the contributions of different clusters to highlight the cluster that dominates in the action features. In this process, by weighted fusion of the impulse signals of different clustering clusters, a unified impulse sequence S f (t) is obtained, thus providing a reliable basis for subsequent feature extraction and action recognition.

[0145] The embodiment of the present invention provides a method for adjusting the weights of an impulse signal set based on a stagger time function and processing the impulse signal set with adjusted weights in a flexible trigger mechanism. Figure 4 The figure shows a schematic implementation flow chart of a method for adjusting the weights of an impulse signal set based on a stagger time function and processing the impulse signal set with adjusted weights in a flexible trigger mechanism. The method for adjusting the weights of an impulse signal set based on a stagger time function and processing the impulse signal set with adjusted weights in a flexible trigger mechanism specifically includes:

[0146] S301, load the impulse signal set, and based on the stagger time function, identify the movement order between multiple skeleton points in the impulse signal set according to the movement time difference between skeleton points.

[0147] S302, capture the dynamic features of the target action, and identify the occurrence of a coherent action by analyzing the time series features between multiple skeleton point categories.

[0148] Among them, the stagger time function τ ij (t) is defined as:

[0149] τ ij (t) = exp(-α|(t j -t i ) - Δt ij |)

[0150] Among them, t i and t j are the timestamps of clusters C i and C j at the t-th moment respectively, α is a parameter controlling the decay of the function, and the deviation of the time series of the clustering center point is exponentially decayed. When the time difference is closer to the preset delay Δt ij , the function value is larger, indicating that the action timing meets the expectation; on the contrary, if the deviation is large, the function value tends to zero, indicating that there is an abnormal timing in the action.

[0151] Output the function τ ij (t) as the dynamic weight and apply it to the impulse signal S f (t):

[0152]

[0153] Among them is the adjusted pulse signal, which dynamically enhances the time difference feature through the staggered time weight τ ij (t).

[0154] It should be noted that the purpose of the staggered time function is to judge the occurrence of a specific action or action sequence by analyzing the time series differences of different skeleton points and their categories during the action. Based on the movement time differences between skeleton points, this function identifies the movement order among multiple skeleton points, thereby accurately capturing the dynamic features of the action and making corresponding judgments. At the same time, by analyzing the time series features among multiple skeleton point categories, the occurrence of coherent actions can be effectively identified.

[0155] S303. Load the set of pulse signals processed by the staggered time function, adaptively extract the features of the pulse signal set in combination with the flexible trigger mechanism based on the water tank, and output the pulse feature set.

[0156] In this embodiment, as Figure 5 shown, a logical schematic diagram of the flexible trigger mechanism based on the water tank is shown. The flexible trigger mechanism based on the water tank processes the set of pulse signals using a nested water tank model;

[0157] When the nested water tank model processes the set of pulse signals, a small water tank state function Q k (t) is defined. The water tank state function Q k (t) represents the cumulative value of the small water tank at time t. The cumulative rule of the water tank state function Q k (t) is as follows:

[0158]

[0159] where represents the input of the pulse signal from the adjusted clustering cluster C k after being adjusted by the staggered time function, and λ t is the linear time decay coefficient of the water tank, which is used to control the automatic emptying speed of the water tank;

[0160] To ensure that the data in the small water tank represents a complete action, a cumulative threshold η k is defined. When Q k (t) ≥ η k , it is considered that the action is completed and the feature is output:

[0161]

[0162] After the feature is output, the small water tank will be reset:

[0163] Q k (t) = 0

[0164] If there is no sufficient input of pulse signals within a period of time, the cumulative value in the small water tank will gradually decrease over time to avoid mis-extracting incomplete action features.

[0165] The nesting mechanism of the nested water tank model is expressed as:

[0166] Multiple small water tanks are nested within a large water tank, and the state function Q D (t) cumulative rule is:

[0167]

[0168] Where, I(Q k (t) ≥ η k ) is the output marking function of the small water tank. When the small water tank outputs a complete action, the value is 1, otherwise it is 0; γ t is the linear attenuation coefficient of the large water tank, and the cumulative threshold of the large water tank is η D , when Q D (t) ≥ η D , it is considered that the whole set of actions is completed, the whole set of action features is output, and the large water tank is reset:

[0169] Q D (t) = 0

[0170] When the cumulative value of the water tank reaches the threshold, the output feature can be expressed as the overall feature F(t) of the serialized pulses of the skeleton points:

[0171] F(t) = {S output (t) | Q D (t) ≥ η D}

[0172] Where, F(t) represents the feature output at time t, and S output (t) is the serialized pulse signal processed by the stagger time function and the water tank mechanism. This mechanism ensures the timing and integrity of feature extraction through the adjustment of the stagger time function and the design of the nested water tank model, effectively improving the accuracy and robustness of action recognition.

[0173] It should be noted that the embodiments of the present invention adopt a staggered time function and a flexible trigger mechanism based on a water tank to achieve adaptive extraction and recognition of features. The staggered time function analyzes the time features in the bone point sequence to separate the pulse signals of different categories, so as to judge the occurrence of actions according to the timing relationship of bone points. In addition, the time series features between multiple bone point categories are also analyzed to identify the occurrence of a coherent action. The flexible trigger mechanism based on the water tank realizes accurate recognition of actions by managing the accumulation and judgment of pulse signals. For example, there is a time difference between the motion peaks of each bone point during the punching process, and the staggered time function can coordinate these time differences to make feature extraction more accurate. The flexible trigger mechanism based on the water tank uses a water tank model to manage the accumulation and judgment of pulse signals. The small water tank represents the determination of a single action. When a set of serialized pulse signals accumulates to the threshold of the small water tank, it means that the action is completed, and its accumulated value is added to the large water tank. The large water tank represents the entire action set. When the data in the small water tanks accumulates to a certain threshold, the features of a complete set of actions are output. The water tank also has a time decay mechanism. If there are not enough pulse signals input within a certain time, the data in the water tank will gradually be emptied to avoid outputting incomplete actions.

[0174] Exemplarily, in a police martial arts combat training scenario, at the beginning of the training, the trainer trains according to the order of police martial arts combat actions. Through the process of extracting bone points, the algorithm identifies and numbers each key bone point, clusters them, and records the coordinates after filtering. Next, pulse conversion and serialization processing are performed on the original data based on the serialized pulse function, and the motion data is converted into a series of pulse signals. These pulse signals are then input into the feature extraction water tank of the system. The water tank mechanism adopts a nested design of a small water tank and a large water tank to gradually accumulate the action features of the trainer:

[0175] (1) Small water tank: First, judge the completion of a single combat action (such as basic punching techniques, kicking techniques, throwing techniques, grappling techniques, etc.).

[0176] (2) Large water tank: Accumulate the actions of multiple small water tanks into the large water tank and calibrate them as complete combined techniques (such as a series of defense and control combined techniques).

[0177] In this embodiment, as Figure 6 shown, the recognition flowchart of the multi-dimensional time series evaluation model is shown. The multi-dimensional time series evaluation model is a pulsed neural network, which consists of an input layer, multiple hidden layers, a pooling layer, and an output layer. The input layer of the network receives the overall feature F(t) of the pulse feature set;

[0178] Among them, the input layer receives the overall feature F(t) as input, and is used to capture the motion pattern of bone points in the time series;

[0179] Each hidden layer corresponds to a specific skeletal point category. The hidden layer processes information through a spiking neuron model. Using the spiking neuron model, the membrane potential u of each neuron i (t) accumulates according to the input features, and the update rule is:

[0180] u i (t) = u i (t - 1) + w i F(t) - τu i (t - 1)

[0181] where w i is the connection weight, and τ is the time constant that controls the decay of the membrane potential.

[0182] The spiking mechanism of the neurons in the multi-dimensional time series evaluation model is as follows:

[0183] When the membrane potential u i (t) exceeds the threshold θ i , the neuron will output a spike signal, and the output signal is defined as:

[0184]

[0185] Here, the output signal o(t) represents the spiking state of the neuron at time t and is used to reflect the occurrence of the motion behavior;

[0186] Between the hidden layer and the output layer, a pooling layer is set up. The pooling layer is used to integrate the output features from multiple hidden neurons and reduce the dimension of the data. The pooling layer uses max pooling to select the maximum value from the outputs of multiple spiking neurons to ensure that the most significant features are focused on in action recognition;

[0187] The pooling operation of the pooling layer is defined as:

[0188] P(t) = max(o1(t), o2(t),..., o n (t))

[0189] where P(t) is the output of the pooling layer, and o n (t) is the spike output of the nth neuron in the hidden layer;

[0190] The output layer is responsible for converting the spike signal of the hidden layer into the final action recognition result, and the calculation method of the output feature is:

[0191]

[0192] where w oiwhere \(w\) is the connection weight between the output layer neurons and the hidden layer neurons, and \(N\) is the number of hidden layer neurons.

[0193] When training the multi-dimensional time series evaluation model, the cross-entropy loss function is used to optimize the performance of the spiking neural network. The loss function of the multi-dimensional time series evaluation model is defined as:

[0194]

[0195] Adjust the connection weight \(w\) in the network through the backpropagation algorithm oi to minimize the loss.

[0196] On the other hand, the embodiment of the present invention also provides an action evaluation system based on staggered time pulses. Figure 7 FIG. shows a schematic structural diagram of an action evaluation system based on staggered time pulses. The action evaluation system based on staggered time pulses specifically includes:

[0197] A data processing module 100, configured to obtain a target action video, extract key skeleton points from the target action video, number the key skeleton points, and cluster the correspondingly encoded skeleton points based on regional anatomical features;

[0198] A pulse processing module 200, which converts serialized motion data into a set of discrete pulse signals through a pulse function and performs serialization processing on the set of pulse signals;

[0199] A feature extraction module 300, configured to load the set of pulse signals, adjust the weights of the set of pulse signals based on the staggered time function, and process the set of pulse signals with adjusted weights in a flexible triggering mechanism to obtain a set of pulse features;

[0200] An action evaluation module 400, configured to construct and train a multi-dimensional time series evaluation model, use the set of pulse features as input, execute the multi-dimensional time series evaluation model, and the multi-dimensional time series evaluation model performs police martial arts fighting action recognition and training evaluation on the target action video, and outputs the action recognition and training evaluation results of the target action video.

[0201] In this embodiment, the data processing module 100, the pulse processing module 200, the feature extraction module 300, and the action evaluation module 400 can be communicatively connected by 5G, DTU, or a local area network.

[0202] Meanwhile, to ensure the reliability of the output features, the action evaluation system based on the off-peak time pulse designed an overtime determination and automatic clearing mechanism to avoid mislabeling of incomplete or interfering action data. The system will evaluate the validity of the accumulated pulse signals within a preset time window. If the action threshold is not reached, the relevant data will be automatically cleared, thereby filtering out noise and interference and improving the stability and accuracy of action labeling.

[0203] On the other hand, an embodiment of the present invention also provides a computer-readable storage medium storing computer program instructions executable by a processor. When the computer program instructions are executed, the methods of any of the above embodiments are implemented.

[0204] As a non-volatile computer-readable storage medium, the memory can be used to store non-volatile software programs, non-volatile computer-executable programs, and modules, such as the program instructions / modules corresponding to the police martial arts combat action evaluation method based on off-peak time pulses in the embodiments of the present application. The memory can include a program storage area and a data storage area. Among them, the program storage area can store an operating system and application programs required for at least one function; the data storage area can store data created for the use of the police martial arts combat action evaluation method based on off-peak time pulses. In some embodiments, the memory may optionally include a memory remotely provided relative to the processor, and these remote memories can be connected to the local module through a network. Examples of the above networks include, but are not limited to, the Internet, enterprise intranets, local area networks, mobile communication networks, and combinations thereof.

[0205] As another aspect of the present invention, a computer device is also provided. The computer device includes a memory and a processor, and a computer program is stored in the memory. When the computer program is executed by the processor, the methods of any of the above embodiments are implemented.

[0206] The various exemplary logic blocks, modules, and circuits described in connection with the disclosure herein can be implemented or executed using the following components designed to perform the functions herein: a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic, discrete hardware components, or any combination of these components. The general-purpose processor can be a microprocessor, but alternatively, the processor can be any conventional processor, controller, microcontroller, or state machine. The processor can also be implemented as a combination of computing devices, such as a combination of a DSP and a microprocessor, multiple microprocessors, one or more microprocessors in combination with a DSP, and / or any other such configuration.

[0207] In summary, the present invention provides a method and system for evaluating police martial arts fighting actions based on staggered time pulses. In the embodiments of the present invention, by introducing a staggered time function to analyze the timing relationship of skeletal points, the actions and continuous actions are judged according to the time series differences of different skeletal points and their categories. The feature adaptive extraction algorithm based on the flexible trigger mechanism of the water tank ensures the timing and integrity of feature extraction. By constructing and training a multi-dimensional time series evaluation model based on a pulse neural network, the cross-entropy loss function is used for training in the design of the multi-dimensional time series evaluation model, so as to optimize the performance of the pulse neural network, improve the accuracy and robustness of motion behavior evaluation, and thus provide a scientific and effective evaluation tool for action training and evaluation, greatly improving the practicality and scientificity of training evaluation. The lightweight design of the pulse neural network enables the system to still process real-time data and recognize actions at a high speed and accuracy under the condition of limited hardware resources. It also has a strong temporal evaluation ability, can monitor and analyze action dynamics in real time, especially in a training environment with high dynamic changes. It overcomes the problems that the existing methods cannot capture the subtle dynamic changes of joints in time and cannot provide a finer-grained motion analysis when dealing with complex motion sequences, and at the same time solves the problems of inefficiency, lack of robustness, adaptability, accuracy and stability of the existing methods. Moreover, the high-dimensional characteristics of human skeletal point data make it difficult for the existing technologies to balance data dimensions and computational efficiency when extracting features.

[0208] The present invention can accurately identify the motion feature categories of different types by capturing the time delay relationship of skeletal point motions through the staggered time function and combining serialized pulse processing. By clustering, the relevant skeletal points are divided into different clusters, thus accurately reflecting their common features during the motion process, ensuring the integrity of feature extraction, and laying a foundation for high-precision action annotation and analysis.

[0209] Through the clustering analysis of skeletal points, the present invention supports the hierarchical management of complex fighting actions. Each clustering center point represents a motion feature. When the pulse signals of multiple clustering center points accumulate to reach a preset threshold, the system calibrates it as a complete set of action moves. This hierarchical accumulation and output mechanism ensures the integrity and structurality of the action sequence, providing strong support for subsequent complex behavior evaluation.

[0210] The present invention converts the original skeletal point data into a pulse signal sequence with significant discrimination through the flexible trigger mechanism based on the water tank, effectively improving the efficiency of feature extraction. The lightweight design of the multi-dimensional time series evaluation model enables the system to still process real-time data and recognize actions at a high speed and accuracy under the condition of limited hardware resources.

[0211] It should be noted that, for the foregoing embodiments, for the sake of simple description, they are all expressed as a series of action combinations. However, those skilled in the art should know that the present invention is not limited by the described action sequence, because according to the present invention, certain steps may be performed in other sequences or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the present invention.

[0212] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than limiting the protection scope of the invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all embodiments. Based on these embodiments, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention. Although the present invention has been described in detail with reference to the above embodiments, those of ordinary skill in the art can still, without conflict and without creative efforts, combine, add, delete or make other adjustments to the features in the embodiments of the present invention according to the circumstances, so as to obtain different technical solutions that are essentially not divorced from the concept of the present invention, and these technical solutions also belong to the scope of protection of the present invention.

Claims

1. A method for evaluating police martial arts fighting movements based on off-peak time pulses, characterized in that, The police martial arts fighting action evaluation method based on staggered time pulses includes: Obtain the target action video, extract the key skeletal points from the target action video, number the key skeletal points, and cluster the correspondingly encoded skeletal points based on the regional anatomical features; Obtain the serialized clustering center point sequence, convert the serialized motion data into a discrete pulse signal set through a pulse function, and perform serialization processing on the pulse signal set; Load the pulse signal set, adjust the weights of the pulse signal set based on the staggered time function, and put the pulse signal set with adjusted weights into a flexible triggering mechanism for processing to obtain a pulse feature set; Construct and train a multi-dimensional time series evaluation model, use the pulse feature set as the input, execute the multi-dimensional time series evaluation model, and the multi-dimensional time series evaluation model performs police martial arts fighting action recognition and training evaluation on the target action video, and outputs the action recognition and training evaluation results of the target action video.

2. The method for evaluating police martial arts fighting movements based on off-peak time pulses as claimed in claim 1, wherein: The method for extracting key skeletal points from the target action video includes: Load at least one set of target action videos, and convert the target action videos frame by frame into at least one set of action images; Use the human pose estimation algorithm BlazePose to detect key skeletal points for each frame of the action image, detect the key skeletal points of the human body, and number the key skeletal points. Each detected skeletal point is assigned a fixed number i; Among them, when the human pose estimation algorithm BlazePose detects key skeletal points for each frame of the action image, it analyzes the human body contour and joint points in the image, estimates the coordinates of each skeletal point in three-dimensional space. Assume that at time t, the coordinates of the skeletal point numbered i are: p i p(t) = (x i (t), y i (t), z i (t)) Among them, (x i (t), y i (t), z i (t)) represents the coordinates of the bone point i in the three-dimensional space; Obtain the detected key skeletal points of the human body, and perform filtering and noise reduction processing on the detected key skeletal points of the human body based on a low-pass filter; Filtered coordinates Calculated by the following formula: When clustering the correspondingly encoded skeletal points based on the regional anatomical features, the clustering result is expressed as: Among them, each cluster C k represents a set of related skeletal points, reflecting the common features existing among them during movement.

3. The method for evaluating police martial arts fighting movements based on off-peak time pulses according to claim 2, characterized in that: The method for converting serialized motion data into a discrete pulse signal set through a pulse function includes: Load the sequence c of cluster centers k (t), calculate the velocity and perform pulse processing on the sequence of cluster centers, and extract the temporal features of the action; Let v k (t) be the velocity of the clustering center point c k (t) at time t, which is defined as: Among them, Δt is the time interval between adjacent sampling times; Preset a speed threshold function θ k , which is used to generate a pulse signal; When the velocity of the center point of the clustering exceeds the preset velocity threshold function θ k it is considered to be in an active state at that moment, and a corresponding pulse signal s k (t) is generated: where s k (t) is the pulse signal of the k-th cluster at time t; Integrate at least one set of pulse signals s k (t) to obtain a discrete set of pulse signals; Load the pulse signal set, perform serialization processing on the pulse signals of each group of clustering center points to obtain a pulse signal in the form of a time series: S k = {s k (1), s k (2), …, s k (T)} Where T is the time length of the entire sequence; Define the weighted fusion pulse signal S f (t): Among them, w k represents the weight of the k-th cluster, which is used to balance the contributions of different clusters to highlight the cluster that dominates the action features. In this process, a unified pulse sequence S f (t) representing the overall action is obtained by weighted fusion of the pulse signals of different clustering clusters.

4. The method for evaluating police martial arts fighting actions based on off-peak time pulses according to claim 3, wherein: The method for adjusting the weights of the pulse signal set based on the staggered time function and putting the pulse signal set with adjusted weights into a flexible triggering mechanism for processing includes: Load the pulse signal set, based on the staggered time function, identify the movement order between multiple skeletal points in the pulse signal set according to the movement time difference between the skeletal points, capture the dynamic features of the target action, and identify the occurrence of coherent actions through the analysis of the time series features between multiple skeletal point categories; Among them, the off-peak time function τ ij (t) is defined as: τ ij (t) = exp(-α|(t j - t i ) - Δt ij |) where t i and t j are the timestamps of clusters C i and C j at the t-th moment respectively, α is a parameter that controls the decay of the function, which exponentially decays the deviation of the time series of the cluster center points. When the time difference is closer to the preset delay Δt ij the larger the function value, indicating that the action timing meets the expectation; conversely, if the deviation is large, the function value tends to zero, indicating that there is an abnormal timing of the action. Output the function τ ij (t) as a dynamic weight and apply it to the pulse signal S f (t): Among them is the adjusted pulse signal, which dynamically enhances the time difference feature through the time staggering weight τ ij (t). Pulse signal set processed by the loading peak shifting time function Combined with the flexible trigger mechanism based on the water tank, adaptively extract the characteristics of the pulse signal set and output the pulse feature set.

5. The method for evaluating police martial arts fighting actions based on off-peak time pulses according to claim 4, characterized in that: The flexible triggering mechanism based on the water tank uses a nested water tank model to process the pulse signal set; When the nested water tank model processes a pulse signal set, a small water tank state function Q k (t) is defined. The water tank state function Q k (t) represents the cumulative value of the small water tank at time t. The cumulative rule of the water tank state function Q k (t) is as follows: Among them, represents the input of the pulse signal from the adjusted clustering cluster C after being adjusted by the staggered time function k of which λ t is the linear time decay coefficient of the water tank, which is used to control the automatic emptying speed of the water tank; To ensure that the data in the small water tank represents a complete action, a cumulative threshold η is defined k , when Q k (t) ≥ η k , it is considered that the action is completed and the feature is output: After the feature is output, the small water tank will be reset: Q k (t) = 0 If there is not enough pulse signal input within a period of time, the accumulated value in the small water tank will gradually decrease with time.

6. The method for evaluating police martial arts fighting actions based on off-peak time pulses according to claim 5, characterized in that: The nesting mechanism of the nested water tank model is expressed as: Multiple small water tanks are nested within a large water tank, and the state function Q D (t) accumulation rule is as follows: where, I(Q k (t) ≥ η k ) is the output marking function of the small water tank. When the small water tank outputs a complete action, the value is 1; otherwise, it is 0. γ t is the linear attenuation coefficient of the large water tank. The cumulative threshold of the large water tank is η D . When Q D (t) ≥ η D , it is considered that the entire set of actions is completed, the characteristics of the entire set of actions are output, and the large water tank is reset: When the cumulative value of the water tank reaches the threshold, the output feature can be expressed as the overall feature F(t) of the serialized pulses of the skeletal points: F(t) = {S output (t) | Q D (t) ≥ η D} Among them, F(t) represents the feature output at time t, and S output (t) is the serialized pulse signal processed by the peak-shifting time function and the water tank mechanism.

7. The method for evaluating police martial arts fighting actions based on off-peak time pulses according to claim 2, characterized in that: The multi-dimensional time series evaluation model is a spiking neural network, which consists of an input layer, multiple hidden layers, a pooling layer, and an output layer. The input layer of the network receives the overall feature F(t) of the spiking feature set; Among them, the input layer receives the overall feature F(t) as input, and is used to capture the motion pattern of the skeletal points in the time series; Each hidden layer corresponds to a specific skeletal point category. The hidden layer processes information through a spiking neuron model. Using the spiking neuron model, the membrane potential u i (t) is accumulated according to the input features, and the update rule is: u i u(t) = i u(t - 1)+w i F(t)-τu i (t - 1) where w i is the connection weight and τ is the time constant that controls the decay of the membrane potential.

8. The method for evaluating police martial arts fighting actions based on off-peak time pulses according to claim 7, wherein: The neuron spiking mechanism of the multi-dimensional time series evaluation model is: When the membrane potential u i (t) exceeds the threshold θ i , the neuron will output a pulse signal, and the output signal is defined as: Here, the output signal o(t) represents the spiking state of the neuron at time t, and is used to reflect the occurrence of the motion behavior; Between the hidden layer and the output layer, a pooling layer is set. The pooling layer is used to integrate the output features from multiple hidden neurons and reduce the dimension of the data. The pooling layer uses max pooling to select the maximum value from the outputs of multiple spiking neurons to ensure that the most significant features are focused on in action recognition; The pooling operation of the pooling layer is defined as: P(t) = max(o1(t), o2(t),..., o n (t)) Among them, P(t) is the output of the pooling layer, and o n (t) is the impulse output of the nth neuron in the hidden layer; The output layer is responsible for converting the pulse signals of the hidden layer into the final action recognition results and outputting features The calculation method is as follows: Among them, w oi is the connection weight between the output layer neurons and the hidden layer neurons, and N is the number of hidden layer neurons.

9. The method for evaluating police martial arts fighting actions based on off-peak time pulses according to claim 7, wherein: When training the multi-dimensional time series evaluation model, a cross-entropy loss function is used to optimize the performance of the spiking neural network. The loss function of the multi-dimensional time series evaluation model is defined as: Adjust the connection weights w in the network through the backpropagation algorithm oi to minimize the loss.

10. An action evaluation system based on staggered time pulses, which is used to implement the method for evaluating police martial arts fighting actions based on staggered time pulses as described in any one of claims 1-9, and is characterized in that: The action evaluation system based on the off-peak time pulses includes: A data processing module, which is used to obtain the target action video, extract the key skeletal points from the target action video, number the key skeletal points, and cluster the correspondingly encoded skeletal points based on the regional anatomical features; A pulse processing module, which converts the serialized motion data into a discrete set of pulse signals through a pulse function and performs serialization processing on the set of pulse signals; A feature extraction module, which is used to load the set of pulse signals, adjust the weights of the set of pulse signals based on the off-peak time function, and put the set of pulse signals with adjusted weights into a flexible triggering mechanism for processing to obtain a set of pulse features; An action evaluation module, which is used to construct and train a multi-dimensional time series evaluation model, use the set of pulse features as input, execute the multi-dimensional time series evaluation model, and the multi-dimensional time series evaluation model performs police martial arts fighting action recognition and training evaluation on the target action video, and outputs the action recognition and training evaluation results of the target action video.

Images