Martial arts kickboxing action recognition training method and system based on depth vision

Abstract

The application discloses a martial arts free combat action recognition training method and system based on deep vision, and relates to the technical field of computer vision and artificial intelligence. The method comprises the following steps: acquiring time sequence three-dimensional skeleton coordinate data sequence of a trainer; generating a starting trend data stream and a trajectory continuity data stream through parallel calculation, performing correlation analysis to generate a behavior starting judgment data stream, and performing joint detection of direction and speed on the trajectory continuity data stream to generate an action interruption data stream; based on logical operation of the behavior starting and action interruption data streams, a behavior authenticity judgment data stream is generated; according to a real attack judgment result, an attack end frame is positioned, posture structure features are extracted as end posture feature data, and a posture reconstruction time length from the attack end to the recovery of defense is calculated to input a pre-trained behavior stage judgment model to output a continuous attack and defense behavior stage data stream. The application realizes automatic and refined recognition of tactical intention and continuous attack and defense stages in martial arts free combat confrontation.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and artificial intelligence, specifically to a method and system for training martial arts and Sanda (Chinese kickboxing) movements based on depth vision. Background Technology

[0002] Wushu Sanda (Chinese kickboxing) movement pattern recognition technology is a research hotspot at the intersection of sports science and computer vision. It aims to capture and analyze athletes' offensive and defensive behaviors through automated means, providing data support for tactical evaluation and combat training. In recent years, with the maturity of depth cameras and skeletal point extraction algorithms, movement classification schemes based on human 3D skeletal sequences have become increasingly intelligent, enabling real-time monitoring of routine technical movements.

[0003] Existing technologies mostly extract spatiotemporal trajectory features of human joints and combine them with deep learning models for semantic labeling to achieve discrete recognition of movement types such as punches and kicks. These methods can demonstrate high recognition accuracy in standardized routine demonstrations and can provide basic normative evaluations based on limb extension range, possessing a certain degree of motion capture capability. However, in highly competitive martial arts combat scenarios, trainees often use a large number of "feints" or "false movements" to mislead their opponents. This is essentially a dynamic interference process in which the limb movement trend deviates from the actual trajectory. This complex transformation of movement intention leads to inconsistencies between the initiation trend of the body's center of gravity and the continuity of the trajectory of the limb extremities.

[0004] Existing technologies often focus only on the morphological matching of limb extremity trajectories, failing to delve into the deep temporal logical connections between the overall body movement trend and the development of local trajectories. This limitation makes it difficult for the system to accurately distinguish deceptive movements that appear to "start" but are actually "interrupted," and it cannot establish a complete offensive and defensive logical loop from the emergence of intent to actual execution, and then to defensive reconstruction after the attack. This not only easily leads to frequent false alarms in the judgment of true and false movements, but also makes it impossible for the system to quantitatively evaluate the athlete's postural stability and defensive transition timing after a real attack, increasing the uncertainty of tactical data statistics in actual combat review, and failing to meet the actual combat needs of Wushu Sanda for refined analysis of dynamic offensive and defensive phases. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method and system for martial arts and combat movement recognition training based on depth vision.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows:

[0007] In a first aspect, the present invention discloses a martial arts Sanda (Chinese kickboxing) movement recognition training method based on depth vision, comprising the following steps:

[0008] Acquire the 3D skeletal coordinate data sequence of the trainee within a continuous temporal sequence;

[0009] Based on the three-dimensional skeletal coordinate data sequence, a starting trend data stream describing the overall movement trend of the body is calculated and generated, and a trajectory continuity data stream describing the consistency of movement at the limb ends is calculated and generated.

[0010] A correlation analysis is performed between the initiation trend data stream and the trajectory continuity data stream to generate a behavior initiation determination data stream;

[0011] The trajectory continuity data stream is subjected to joint detection of direction and velocity to generate an action interruption data stream;

[0012] Logical operations are performed on the behavior initiation determination data stream and the action interruption data stream to generate a behavior authenticity determination data stream;

[0013] Based on the determination results representing real attacks in the behavior authenticity determination data stream, the attack end frame is located in the three-dimensional skeleton coordinate data sequence, and the posture structure features of the attack end frame are extracted to generate end posture feature data.

[0014] Calculate the time interval from the end of the attack frame to reaching the preset defensive posture, and generate posture reconstruction time data;

[0015] The behavior authenticity determination data stream, the ending posture feature data, and the posture reconstruction duration data are input into the pre-trained behavior stage determination model, and the attack and defense behavior stage data stream corresponding to the time sequence is output.

[0016] Secondly, this invention discloses a martial arts Sanda (Chinese kickboxing) movement recognition training method system based on depth vision. The method, using the aforementioned depth vision-based martial arts Sanda movement recognition training method, includes:

[0017] The data acquisition module is used to acquire the three-dimensional skeletal coordinate data sequence of the trainee in a continuous time sequence;

[0018] The data processing module is used to calculate and generate a starting trend data stream describing the overall movement trend of the body based on the three-dimensional skeletal coordinate data sequence, and to calculate and generate a trajectory continuity data stream describing the consistency of movement at the limb ends.

[0019] The correlation analysis module is used to perform correlation analysis on the initiation trend data stream and the trajectory continuity data stream to generate a behavior initiation judgment data stream;

[0020] The dynamic extraction module is used to perform joint detection of direction and velocity on the trajectory continuity data stream to generate an action interruption data stream;

[0021] The intent determination module is used to perform logical operations on the behavior start determination data stream and the action interruption data stream to generate a behavior authenticity determination data stream.

[0022] The attack and defense assessment module is used to locate the attack end frame in the three-dimensional skeleton coordinate data sequence based on the judgment result representing the real attack in the behavior authenticity judgment data stream, and extract the posture structure features of the attack end frame to generate end posture feature data.

[0023] The attitude reconstruction module is used to calculate the time interval from the end of the attack frame to the achievement of the preset defensive attitude, and generate attitude reconstruction time data.

[0024] The phase fusion module is used to input the behavior authenticity determination data stream, the end posture feature data, and the posture reconstruction duration data into the pre-trained behavior phase determination model, and output the attack and defense behavior phase data stream corresponding to the time sequence.

[0025] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0026] 1. This invention effectively filters out motion interference from non-real attacks by calculating the consistency between the overall body movement trend and the movement of the extremities and performing correlation analysis. It captures the explosive displacement of the center of gravity using a startup trend data stream and verifies the execution depth of the extremities by combining it with a trajectory continuity data stream, enabling the system to establish a temporal logical relationship between intent germination and trajectory development. This correlation analysis method based on multi-source motion information overcomes the inaccuracies caused by relying solely on limb geometry in traditional visual recognition, thus ensuring the objectivity and accuracy of determining real attack behavior.

[0027] 2. This invention establishes a technical mechanism for identifying deceptive actions. By monitoring the direction reversal and speed changes of the trajectory in real time, the system can identify unnatural deceleration or trajectory truncation phenomena that occur after effective activation, thereby accurately separating deceptive actions from the real attack sequence. This process allows action recognition to go beyond simple waveform matching and delve into the coherence determination of the action execution process, thus providing more realistic intent recognition feedback for practical training.

[0028] 3. This invention achieves a quantitative evaluation of the effectiveness of attack-defense transition by locating the attack termination frame and extracting posture structure features, combined with the calculation of posture reconstruction time. By integrating behavioral realism with the body stability, head deviation, and center of gravity at the moment of termination, the system can quantitatively describe the dynamic process from the completion of the attack action to the reshaping of the defensive posture. Attached Figure Description

[0029] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings, the same reference numerals are used to refer to the same parts. Wherein:

[0030] Figure 1 This is a flowchart of the steps of the present invention;

[0031] Figure 2 This is a schematic diagram illustrating the working principle of the present invention;

[0032] Figure 3 This is a system module connection diagram of the present invention. Detailed Implementation

[0033] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0034] In existing technologies, martial arts and Sanda (Chinese kickboxing) movement recognition largely relies on single-body skeletal trajectory matching or end-to-end deep learning classification, making it difficult to balance the authenticity of movement intent recognition with the detail of stage segmentation. Traditional methods, when dealing with common feints in real combat, lack logical verification of the consistency between the center of gravity initiation and the final trajectory, easily misinterpreting feints as real attacks, resulting in severely compromised recognition accuracy due to dynamic interference. Existing systems cannot simultaneously perceive the constraints of body center of gravity trends on limb force exertion, especially during fast-paced combat when movements are interrupted or connected; static models exhibit discrimination lag and logical gaps, failing to meet the needs of quantitative analysis of Sanda techniques and tactics.

[0035] To address the aforementioned issues, the study discovered a spatiotemporal correlation between the acceleration of the body's center of gravity displacement and the directional consistency of the limb's end-effector trajectory. A coupled model of initiation trend and trajectory continuity was established to differentiate between genuine and feigned movements. Further findings revealed that center of gravity acceleration is highly sensitive to force application intent, while end-effector trajectory consistency provides strong support for the stability of movement execution. Therefore, a method was proposed to dynamically determine the authenticity of behavior based on initiation judgment data and interruption detection logic. Further experimental verification incorporated the mapping relationship between behavior authenticity judgment and posture reconstruction duration into the parameter fusion mechanism of the behavior phase judgment model, forming a closed-loop feedback system covering preparation, execution, recovery, and defensive reconstruction.

[0036] After introducing the basic concept of the present invention, the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0037] Example 1:

[0038] like Figure 1 As shown, the training method for martial arts Sanda movement recognition based on depth vision includes the following steps:

[0039] Acquire the 3D skeletal coordinate data sequence of the trainee within a continuous temporal sequence;

[0040] Based on the three-dimensional skeletal coordinate data sequence, a starting trend data stream describing the overall movement trend of the body is calculated and generated, as well as a trajectory continuity data stream describing the consistency of movement at the limb ends is calculated and generated.

[0041] A correlation analysis is performed between the initiation trend data stream and the trajectory continuity data stream to generate a behavior initiation judgment data stream;

[0042] Joint detection of direction and velocity is performed on the trajectory continuity data stream to generate an action interruption data stream;

[0043] Logical operations are performed on the behavior initiation determination data stream and the action interruption data stream to generate the behavior authenticity determination data stream;

[0044] Based on the judgment results representing real attacks in the behavior authenticity judgment data stream, the attack end frame is located in the three-dimensional skeleton coordinate data sequence, and the posture structure features of the attack end frame are extracted to generate end posture feature data.

[0045] Calculate the time interval from the end of the attack frame to reaching the preset defensive posture, and generate posture reconstruction time data;

[0046] Input the behavior authenticity determination data stream, the ending posture feature data, and the posture reconstruction duration data into the pre-trained behavior stage determination model, and output the attack and defense behavior stage data stream corresponding to the time sequence.

[0047] like Figure 2 The diagram shown illustrates the workflow of this application. The core working principle of this application is built upon a highly collaborative foundation of underlying hardware sensing and upper-level logic computation. During the data acquisition phase, the system uses 3D vision sensing hardware (such as depth cameras, infrared optical capture arrays, or lidar sensing systems) deployed around the training area to scan the trainee's movement space in real time. The hardware acquisition unit continuously captures the spatial radiation information of human joints at a preset frequency (e.g., 30Hz-120Hz), which is then processed by an embedded processing chip using depth map conversion and skeleton extraction algorithms. The final output is a sequence of 3D skeletal coordinate data composed of multi-dimensional vectors.

[0048] The system processes the 3D skeletal coordinate data sequence in parallel. First, by extracting the displacement changes of key points or center of gravity points on the human torso, it calculates the instantaneous vector direction and magnitude of the body's center of mass in 3D space, thereby generating a starting trend data stream describing the overall movement trend of the body. Second, focusing on the attack extremities such as hands, feet, elbows, and knees, it calculates the slope of the trajectory tangent and the magnitude of the motion vector at each extremity joint between adjacent frames, generating a trajectory continuity data stream describing the consistency of limb extremity movement. Subsequently, the system executes correlation analysis instructions, performing time alignment and spatial correlation calculations on the center of gravity burst signal in the starting trend data stream and the end-point direction in the trajectory continuity data stream. By determining whether the center of gravity displacement and the end-point trajectory achieve coordination within the preset physical logic, it generates a behavior initiation judgment data stream.

[0049] To address unrealistic interference during the attack, the system jointly monitors the directional consistency and instantaneous velocity modulus of the trajectory continuity data stream. By detecting the smoothness of the trajectory waveform and the presence of unnatural kinetic energy decay, it identifies interruptions, withdrawals, or feints, generating an action interruption data stream. The system performs Boolean or probabilistic logic operations based on the action initiation determination data stream and the action interruption data stream. The action interruption data stream reflects inconsistencies in the attack execution process, such as sudden drops in speed or reversals in direction. The system aligns this data stream with the action initiation determination data stream in sequence and uses logical operations to determine whether the initiation is accompanied by an interruption, thus determining the authenticity of the action. For example, if it is determined to be "initiated" and "not interrupted," the result points to a real attack; conversely, if it is detected to be "initiated" but "interrupted" or "direction deviated," it is determined to be a feint.

[0050] In the offensive and defensive effectiveness evaluation phase, the system performs a backtracking search in the cached temporal 3D skeletal coordinate data sequence based on the judgment signal of the actual attack, accurately locating the attack end frame where the kinetic energy of the movement is released. The system extracts the spatial topology of the whole body skeleton in this frame, including the angles between joints, the body's windward area, and the supporting base area, generating end posture feature data. At the same time, the system monitors the time span required for the trainee's skeletal structure to revert to the system's preset standard defensive stance template from this attack end frame, calculating and generating posture reconstruction time data.

[0051] The system concatenates the feature dimensions of the data stream representing the authenticity of the behavior, the end posture feature data representing the instantaneous defensive state, and the posture reconstruction duration data representing the conversion efficiency, and inputs them into a pre-trained behavior stage determination model. This model, based on a long short-term memory network or temporal convolutional network structure, performs a nonlinear mapping on the input composite data stream, ultimately outputting a data stream of offensive and defensive behavior stages that completely corresponds to the original temporal sequence and covers the entire process from action initiation to defense reconstruction.

[0052] In the pre-training stage of the behavior phase determination model, the system employs an offline deep learning framework optimized for skeletal coordinate datasets converted from real adversarial videos. The network structure utilizes a fusion architecture of a bidirectional long short-term memory network (Bi-LSTM) and a temporal convolutional network (TCN), leveraging a self-attention mechanism to extract cross-dimensional features of realism judgment, ending pose, and reconstruction duration, capturing long-distance dependencies between action phases. The behavior phase determination model employs a three-layer bidirectional LSTM network with 128 neurons per layer, followed by fully connected layers and a softmax output layer. The training dataset contains skeletal sequences extracted from 500 hours of professional Sanda (Chinese kickboxing) competition videos, covering 10 common attack and defense patterns.

[0053] The model employs the Adam optimizer, with an initial learning rate set between 1e-3 and 1e-4, and incorporates a cosine annealing strategy to dynamically decay the learning rate. The number of training epochs is typically set to 100 to 200, with an early stopping mechanism to prevent overfitting. The batch size is set to 32 or 64 MB depending on the hardware memory configuration. To enhance the model's generalization ability, a packet loss rate (0.3-0.5%) is added to the hidden layers, and data augmentation processing such as translation, rotation, and scaling of the skeletal coordinates is performed to ensure that the model can output stable and accurate behavioral stage data streams when handling different trainees' body types and movement habits.

[0054] This application establishes a spatiotemporal correlation mechanism between body center of gravity trends and end-effector trajectory continuity, enabling deep discrimination of martial arts Sanda (Chinese kickboxing) movement intentions and solving the technical challenge of traditional visual recognition's inability to filter out feints. Through quantitative calculation of the posture structure and reconstruction duration of the attack end frame, the system can objectively describe the timeliness and stability indicators during the offensive-defensive transition process. This not only improves the physical consistency of movement phase divisions but also provides athletes with feedback data based on technical facts, effectively supporting the refinement and objectivity of training evaluation.

[0055] This application further proposes that, during the generation of the trend data stream, the system performs the following specific steps: The system extracts the spatial coordinate sequence of the trainee's torso center point (such as the hip center or center of mass) from the acquired three-dimensional skeletal coordinate data sequence. By subtracting the coordinate values ​​of this point between consecutive frames, the system obtains a displacement vector reflecting the direction and magnitude of the center of gravity's movement in three-dimensional space.

[0056] To further quantify the explosive intensity of a movement, the system performs differential calculations on the displacement vector to obtain an acceleration vector sequence. Specifically, the acceleration vector sequence is derived by calculating the rate of change of the displacement vector over time. In actual combat, the magnitude of the acceleration vector sequence directly corresponds to the intensity of the body's force exertion and is a key physical indicator that distinguishes between everyday movement and the initiation of an attack.

[0057] Simultaneously, the system establishes posture monitoring logic based on the human sagittal plane. By monitoring the projection position of the trunk center point on the sagittal plane (i.e., the lateral plane of the body), the system calculates the rate of change of the trunk center point's projection on the human sagittal plane relative to its previous position (e.g., calculating its circular motion offset relative to the previous moment above the supporting joint (such as the supporting ankle)). This rate of change is used as the equivalent posture change rate of the trunk's rotational tendency around the supporting joint. The equivalent posture change rate describes how fast the trunk rotates around the supporting axis, reflecting the kinematic tendency of the athlete to assist in force generation through forward leaning or twisting of the center of gravity.

[0058] The system generates a fusion value by weighting and fusing the magnitude of the acceleration vector sequence (representing linear explosive force) with the equivalent attitude change rate (representing rotational trend). In specific engineering implementations, the system adopts a simple linear weighting method, that is, by assigning corresponding weight coefficients to the acceleration and velocity terms (such as 0.5 for each, or dynamically adjusting according to the type of punch and kick), a comprehensive fusion value is calculated.

[0059] When the real-time calculated fusion value exceeds the preset fusion threshold, the system immediately determines it as a valid start and generates a start trend marker in the current time series data, which serves as the trigger source for subsequent verification of true and false actions. The preset fusion threshold Y (typically ranging from [3.5, 4.2]) is set by statistically analyzing the fusion value distribution of 50 professional Sanda athletes in real attack actions, taking the 90th percentile, and is exemplarily 3.8.

[0060] By employing dual constraints of acceleration and posture rotation rate, this application achieves digital modeling of the overall body movement trend. Acceleration features ensure sensitive capture of explosive force, while the rate of change of sagittal projection effectively filters out simple forward body movement or unintentional fluctuations in the center of gravity. This multi-index coupled judgment method not only reduces recognition noise caused by single parameters but also provides a rigorous biomechanical basis for the generation of initiation trend markers, improving the response accuracy and reliability of the motion recognition training system in complex adversarial environments.

[0061] This application further proposes specific steps for calculating and generating a trajectory continuity data stream describing the consistency of limb extremity movements. By utilizing the dynamic stability characteristics of microscopic trajectories, the energy penetration of the attack action during execution is characterized.

[0062] During system operation, the system locks a temporal 3D skeletal coordinate data sequence representing the limb extremities of the attacking weapon (such as fist, elbow, ankle, or knee). By calculating the ratio of the spatial displacement between adjacent sampling frames to the sampling period, the instantaneous velocity vector of each extremity joint is obtained. This instantaneous velocity vector not only contains the speed information of the extremity movement but also carries a crucial spatial orientation dimension, serving as the core data foundation for measuring whether the action execution is "distorted".

[0063] To quantify the directional stability of extremities during high-speed motion, the system introduces spatial geometric analysis based on a sliding window mechanism. Within a sliding window of a preset length of M frames, the average cosine similarity between the instantaneous velocity vector in the current frame and the historical frames within the window is calculated as the directional consistency coefficient. The sliding window of the preset length of M frames (typically ranging from [5,15] frames, corresponding to a motion window of approximately 0.1-0.3 seconds, depending on the sampling frequency) dynamically advances along the time axis.

[0064] In one embodiment, within a sliding window of a preset length of M frames, the system extracts the velocity vector of the current frame. velocity vectors of each historical frame within the window The angle between the two frames is calculated, and cosine similarity is used to evaluate the consistency of each frame, resulting in the directional consistency coefficient C. The formula for calculating the directional consistency coefficient C is as follows:

[0065] ;

[0066] The directional consistency coefficient C is in the range of [-1, 1]. The closer its value is to 1, the more the movement direction of the terminal joint tends to be a straight line or a smooth arc in a short period of time. This is an important criterion for judging the continuity of the transmission of real attack force.

[0067] Meanwhile, considering the data noise caused by physiological tremors or minor adjustments in the human body during non-aggression states, the system performs high-pass filtering on the end-effector displacement sequence. The system performs high-pass filtering on the displacement sequence of the joints at the limb ends, filtering out motion components below a preset frequency threshold, and then uses Kalman filtering or a five-point cubic smoothing algorithm to further refine the trajectory, generating smoothed trajectory data. The preset frequency threshold is typically set between [2Hz, 5Hz], the logic being that effective attack movements in martial arts (such as punches or kicks) usually have extremely short burst times and high characteristic frequencies, while spontaneous random swaying of the human body occurs at lower frequencies.

[0068] The system encapsulates the directional consistency coefficient and smoothed trajectory data in a time sequence to generate a trajectory continuity data stream. The pre-training process of the trajectory continuity data stream mainly models the trajectory volatility of athletes at different levels. In a standardized laboratory environment, by comparing the differences in end-effector kinetic energy loss between elite athletes and beginners, an effective benchmark for determining the consistency coefficient is established.

[0069] By correlating the spatial angle of instantaneous velocity vectors and using high-pass filtering for noise reduction, this application achieves a refined representation of the physical quality of attack trajectories. The system can identify trajectory deflections caused by feints or insufficient force during limb movement, while the smoothed trajectory data ensures the purity of displacement calculations. This coupled processing based on spatial geometry and frequency domain filtering ensures that the continuous trajectory data stream can accurately reflect the execution depth of the attack action.

[0070] This application further proposes specific steps for generating a behavior initiation determination data stream by performing correlation analysis on the initiation trend data stream and the trajectory continuity data stream. In the system's operational logic, the system performs feature retrieval on the real-time generated initiation trend data stream, capturing the moment of explosive instability of the body mass center (COM) by identifying the "initiation trend markers" carried within it. The initiation trend markers are defined as candidate reference points for action triggering, providing a precise temporal index for subsequent local feature extraction.

[0071] The system detects the moment when the trend marker is activated. Starting from a certain point, a preset activation association window of length W is opened. The length W of the preset activation association window needs to take into account the transmission delay of human physiological force exertion (i.e., the time difference from the start of the movement from the center of gravity to the force exertion at the extremities). According to statistics from actual combat data, the length W of this preset activation association window is usually set between [150ms, 450ms]. During this window period, the system retrieves the synchronously generated trajectory continuity data stream and performs a two-dimensional judgment on the dynamic performance of the limb extremities.

[0072] In terms of the discrimination mechanism, the system continuously monitors the directional consistency coefficient C within the window. It determines whether the mean or duration of this directional consistency coefficient C must be higher than a preset consistency threshold within the preset activation associated window. (The typical value range is [0.85, 0.92]. It is set by taking the 95th percentile of the distribution of the directional consistency coefficient of effective attack movements of 100 professional athletes, and is exemplarily 0.88). The physical meaning of this condition is that only when the limb ends show a highly stable directional movement in the short time after the center of gravity starts to move can it be considered to have a real attack direction.

[0073] Subsequently, the system extracts the displacement components of the smoothed trajectory data in the principal spatial direction (i.e., the expected attack vector direction) using principal component analysis (PCA) or projective decomposition. Only when the displacement component exceeds the preset displacement threshold. (Based on the statistical average of the effective attack distance in the Sanda competition rules, and combined with the ratio of the athlete's arm span to leg length, the normalization setting is usually set to [15cm, 25cm] for punching actions and [30cm, 50cm] for leg actions. For example, punching is 18cm and leg action is 40cm.) Only when the action is deemed to have sufficient physical aggression.

[0074] The above discrimination logic can be represented by the following decision operator:

[0075] ;

[0076] If all conditions are met simultaneously, a behavior initiation determination flag is generated. This data is then used as the behavior initiation judgment data stream. This logic ensures that the system can exclude invalid actions that, although there is body swaying (generating an initiation trend marker), do not form an effective attack trajectory. The parameter optimization process of the above correlation mechanism is completed by comparing the dynamic differences between "feints" and "real moves" in a large number of Sanda combat videos. In the pre-training phase, the system uses labeled time anchors to perform regression analysis under different window lengths and threshold combinations to determine the parameter boundaries that minimize the false alarm rate.

[0077] Through a serial correlation logic of "overall trend triggering and local trajectory verification," this application achieves the transformation from vague body initiation to clear behavior determination. The continuous requirement of the directional consistency coefficient filters out irregular limb swings, while the hard constraint of spatial main direction displacement ensures the physical effectiveness of the determination marker. This dual threshold constraint mechanism enables the behavior initiation determination data stream to not only have extremely high time synchronization, but also to rigorously define "real force exertion" at the logical level, laying a solid foundation for subsequent differentiation of real and false actions and fine-grained stage recognition.

[0078] This application further proposes specific steps for jointly detecting direction and velocity in a trajectory continuity data stream to generate an action interruption data stream.

[0079] The system calculates the instantaneous acceleration of the end joints in the smoothed trajectory data between consecutive frames by numerically differentiating the instantaneous velocity vectors between adjacent frames. This acceleration, or velocity change rate, reflects the surplus or deficit of limb kinetic energy.

[0080] The system performs critical motion direction reversal detection, and its core logic lies in monitoring and detecting the velocity vectors of adjacent frames of smoothed trajectory data. and The spatial angle changes. Among them, the reversal of the motion direction is determined by the dot product of the velocity vectors of adjacent frames changing from positive to negative.

[0081] The formula for determining the reversal of the direction of motion follows the logic of the dot product P operation, and the formula is as follows:

[0082] ;

[0083] When dot product When the value suddenly changes from positive to negative, it means the velocity angle between the two frames. Exceed (Usually when withdrawing a move, it tends to be close to) The system immediately determined that the direction of motion had reversed.

[0084] An action interruption flag is generated when the rate of change of velocity is detected to be lower than a preset threshold and a reversal of motion direction is simultaneously detected. Specifically, to eliminate interference from natural damping caused by target impact or normal arcing motion, the system establishes a dual-criteria coupled decision mechanism. Only when the rate of change of velocity is detected to be lower than the preset threshold... (That is, a significant drop in kinetic energy occurs. This is determined by analyzing the natural deceleration distribution after a real attack action hits the target under high-speed photography, and combining the statistical differences in the artificial braking acceleration in the feint. The midpoint of the dividing range between the two is taken, with a typical value range of [-15m / s²].) 2 -40m / s 2 The system will only generate an action interruption flag when the above-mentioned motion direction reversal signal is triggered simultaneously (between [the specified points]). This logic ensures that only those behaviors that suddenly stall during high-speed motion and are accompanied by a reversal of trajectory will be marked as "interrupted".

[0085] Furthermore, the system further subdivides the interruption flags based on the physical characteristics of the action failure. This is done based on the steepness of the rate of change of velocity (i.e., the time derivative of acceleration, jerk) and the range of the angle of reversal of the motion direction (e.g.,...). to It was determined to be a complete withdrawal. to If the action is determined to be a mid-course change of direction, a type label (such as "passive blocking", "active withdrawal" or "change of direction feint") is assigned to the action interruption marker, forming a multi-dimensional action interruption data stream.

[0086] The system's discrimination threshold and type mapping relationship originates from cluster analysis of numerous "effective hits" and "active withdrawals" samples from real-world combat scenarios. During pre-training, the model iterates its sensitivity across action sequences at different load levels to ensure the preset change threshold is maintained. The parameters are designed to adapt to the rapid braking of a punch and the high inertia of a leg technique's retraction.

[0087] By linking the positive and negative conversion of the velocity vector dot product with the acceleration threshold, this application achieves accurate deconstruction of discontinuous action intentions. The introduction of the velocity change rate provides a quantitative basis for kinetic energy decay, while the direction reversal detection provides a logical determination of the spatial trajectory. The combination of the two effectively solves the problem of identifying "false start of action" in complex adversarial environments. This closed-loop detection mechanism based on kinematic abrupt change features enables the action interruption data stream to have temporal penetration, providing indispensable subtraction data for the logical calculation of the subsequent behavior authenticity.

[0088] This application further proposes implementation details regarding the logical operations performed on the behavior initiation determination data stream and the action interruption data stream to generate a behavior authenticity determination data stream. The system places the two sets of heterogeneous marker streams generated under a unified time coordinate system for synchronous verification, and strips away false interference information through temporal causality to establish the physical authenticity of the attack behavior.

[0089] The system performs high-precision timing alignment. Because there is a slight processing lag in the computation chain between the behavior initiation determination marker (representing the burst start of an action) and the action termination marker (representing the non-natural termination point of an action), the system needs to remap these two types of marker streams on the timing axis based on the timestamps of the original skeletal sampling frames. This step ensures that logical operations are performed within the same action lifecycle, avoiding erroneous associations across action features.

[0090] The system uses the logical start time of the start-up judgment flag for each behavior. Based on this, a preset authenticity determination window is opened. The length of this window is designed based on the maximum duration of a single full-speed attack action in human biomechanics, typically set between [600ms, 1200ms]. Before the window closes, the system continuously searches for a corresponding action interruption marker. If the preset authenticity determination window is in progress... Within, the logical start time point of the behavior initiation determination flag is detected. If at least one action interruption marker is detected afterward, it indicates that although the limb strike had the characteristics of initiation, its kinetic energy subsequently underwent abnormal decay or trajectory reversal due to human intervention. Based on this, the system determines that the action was initiated as a "fake action". Conversely, if no action interruption marker is detected after the window period expires, and the trajectory continuity is maintained until the kinetic energy is naturally dissipated or a collision occurs, it is determined to be a "real attack".

[0091] This logical operation process can be represented by a Boolean logic discriminant as follows:

[0092] ;

[0093] in, This represents the result of determining the authenticity of the behavior; a value of 1 indicates a real attack, while a value of 0 and the presence of an initiation trend marker classifies it as a fake action.

[0094] This is a marker for movement interruption, indicating that the trajectory of the limb's end has undergone an unnatural reversal of movement direction (the dot product changes from positive to negative) or a sudden drop in instantaneous velocity, indicating an abnormal decay of kinetic energy;

[0095] It is a time variable, representing the timestamp of the current frame in the time sequence, used for logical retrieval in a continuous data stream;

[0096] This is a marker for determining the initiation of a behavior, indicating that the initiation trend of the body's center of gravity and the trajectory of the limbs have been matched, proving that the movement has entered the substantial force exertion stage.

[0097] Finally, the system streams the identified fake action tags and real attack tags according to the original timestamp order of each behavior initiation judgment tag. Each group of tags is associated with its corresponding spatial displacement vector and energy feature, thus forming a structured behavior authenticity judgment data stream.

[0098] Preset authenticity determination window To address logical redundancy, the model was pre-trained and cross-validated using large-scale real-world combat samples. During training, the model's parameters were optimized under extreme conditions such as complex combos and rapid feints to ensure that the decision-making logic, while maintaining real-time performance, could minimize misjudgments of rapid, continuous attacks.

[0099] Through the aforementioned temporal logic architecture, this application effectively addresses the technical challenge of distinguishing between "tactical feints" and "real attacks" in depth vision recognition. This logical operation mechanism does not rely on complex image feature classification but instead returns to the causal chain of the essence of motion, resulting in highly reliable logical data streams for determining the authenticity of actions. This not only provides crucial attribute basis for the accurate division of subsequent offensive and defensive phases but also lays a solid technical foundation for the quantitative assessment of athletes' tactical skills.

[0100] This application further proposes specific steps for extracting the pose structure features of the attack end frame and generating end pose feature data. Regarding the implementation details of extracting the pose structure features of the attack end frame and generating end pose feature data, this solution aims to transform the instantaneous static skeleton layout into a quantitative defense quality evaluation index.

[0101] Once the system determines that a real attack has reached the end of its kinetic energy release, it immediately locks the end frame of the attack and deconstructs the physiological structural attributes of the finishing move of a Sanda (Chinese kickboxing) action by extracting the spatial relationships of key anatomical points in the human body.

[0102] Based on the three-dimensional skeletal coordinate data sequence of the attack end frame, calculate at least one geometric feature value characterizing body stability, at least one distance feature value characterizing head offset, and body center of gravity height value.

[0103] Geometric eigenvalues ​​are typically based on the principle of the supporting polygon, which calculates the spatial offset vector of the center of gravity projection point relative to the geometric center formed by the contact points between the supporting foot (or both feet) and the ground. Additionally, the system calculates the angle between the torso's longitudinal axis and the ground normal to assess the body's balance slope. The calculation logic for the stability geometric eigenvalue S can be expressed by the following formula:

[0104] ;

[0105] in, The area of ​​the supporting surface;

[0106] This is the geometric distance between the center of gravity projection point and the center of the support surface. The larger this value is, the higher the stability of the lower body at the end of the attack.

[0107] The system calculates a distance feature value representing head offset. In Wushu Sanda combat, the spatial position of the head is directly related to defensive safety. The system extracts the head displacement vector and calculates its lateral deviation distance relative to the torso's central axis. If this deviation distance exceeds a preset safety threshold (e.g., [8cm, 15cm]) at the moment the attack ends, it is determined that there is a defensive vulnerability. This feature value reflects whether the trainee can effectively control the head position after exerting force, avoiding unnecessary swaying.

[0108] The system extracts the height of the body's center of gravity in real time. Center of gravity height is a key indicator of dynamic transition flexibility. The system obtains a normalized center of gravity height value by weighted averaging the mass distribution and real-time coordinates of all body parts (such as the trunk, upper limbs, and lower limbs). A lower and more stable center of gravity usually corresponds to better impact resistance and faster second-start capability.

[0109] The system performs feature concatenation and vector normalization on geometric feature values, distance feature values, and centroid height values ​​to generate multi-dimensional end posture feature data. This generation process does not rely on subjective visual impressions but is based on rigorous geometric and mechanical calculations, ensuring the objectivity of the evaluation dimensions.

[0110] The discrimination criterion is derived from data clustering of the finishing postures of high-level Sanda athletes. During pre-training, the behavioral phase judgment model extracts the optimal posture feature range that can maintain balance under simulated impact, serving as a benchmark for measuring the quality of the trainee's posture.

[0111] By integrating stability, offset, and center of gravity height in a three-dimensional calculation, this application achieves precise digital modeling of the "recovery state" after an attack. The introduction of geometric eigenvalues ​​quantifies balance capability, offset distance eigenvalues ​​pinpoint defensive blind spots, and center of gravity height reveals the potential for subsequent actions. This multi-dimensional combination of features enables the end-posture feature data to not only describe the quality of the current action's completion but also possess the technical depth to predict the difficulty of subsequent defensive reconstruction, providing highly instructive spatial state input for overall behavioral phase determination.

[0112] This application further proposes specific steps for calculating the time interval from the end of the attack frame to reaching the preset defensive posture, and generating posture reconstruction time data.

[0113] In practice, the system predefines a standard defensive stance joint coordinate template in the storage unit. This standard stance template is constructed by pre-collecting the spatial distribution of joint coordinates of high-level athletes in standard defensive states (such as the Sanda stance), and is normalized to adapt to trainees of different body types, serving as a physical benchmark for measuring stance regression.

[0114] The system initiates real-time comparison logic starting from the frame where the attack has ended. The system extracts the trainee's 3D skeletal coordinates frame by frame and calculates the similarity score between the skeletal coordinates of the current frame and the standard defensive posture template.

[0115] In engineering implementation, the similarity score is determined by calculating the spatial distance deviation of key joints (such as the fist, elbow, shoulder, and core): the smaller the overall geometric distance between the joint and the template, the higher the score. This direct comparison based on spatial position can objectively reflect the degree to which the current posture approaches the standard defensive state.

[0116] To ensure the stability of the judgment results, the system introduces a temporal window verification mechanism. When the similarity score is higher than a preset similarity threshold for N consecutive frames... When the preset defensive posture is reached, it is determined. Preset similarity threshold. (Typical value range is [0.85, 0.95]). By calculating the variance distribution of joint coordinates of multiple athletes in the standard defensive posture, the mean plus twice the standard deviation is taken as the lower limit for judging the stable posture, which is 0.90 for example.

[0117] Where N is an integer greater than 1. The system not only requires a single-frame score to meet the standard, but also requires the similarity score to be consistently higher than the preset similarity threshold for N consecutive frames (N is an integer greater than 1, for example, 3 to 5 frames depending on the sampling frequency). This setting effectively eliminates misjudgments caused by momentary trajectory overlap or data fluctuations during the execution of the action, ensuring that the trainee has indeed stably controlled the body's return and maintained the defensive posture.

[0118] Once the continuous compliance condition is met, the system immediately subtracts the timestamp of the first frame that indicates compliance from the timestamp of the attack end frame to obtain the time difference between the two, and uses this difference as the attitude reconstruction duration data. In one embodiment, the timestamp of the attack end frame is... The timestamp of the first frame that determined the defensive stance was reached was... Then the attitude reconstruction time data The calculation formula is This duration data intuitively reflects the time it takes for an athlete to recover from an unstable offensive state to a balanced defensive state. This technical solution achieves precise quantification of the rhythm of offensive and defensive transitions through a combination of "benchmark template matching" and "temporal stability filtering." This not only provides objective numerical support for evaluating an athlete's defensive awareness but also enables the training system to provide specific improvement suggestions for the key technical detail of delayed recovery, thereby enhancing the effectiveness of training.

[0119] Through a time-based calculation logic of "template comparison and continuous verification," this application achieves a quantitative evaluation of the "recovery efficiency" of Sanda (Chinese kickboxing) movements. The introduction of standard posture templates provides an objective benchmark for the evaluation system, while the discrimination constraint of N consecutive frames ensures the robustness of state transitions. This duration calculation mechanism is based on spatiotemporal similarity.

[0120] This application further proposes specific steps for outputting attack and defense behavior phase data streams corresponding to time-series sequences. Through spatiotemporal fusion of heterogeneous data, the system transforms combat actions from discrete feature points into a dynamic process description with semantic coherence. The system is responsible for processing information across three dimensions: logical discrimination, spatial topology, and time efficiency, ultimately generating a standardized phase stream that can be used for practical review.

[0121] The system performs channel fusion and temporal normalization of multidimensional data. The behavior authenticity determination data stream (logical label), the ending posture feature data (spatial vector), and the posture reconstruction duration data (time scalar) are physically different, and the system needs to map them to a unified time axis coordinate system. The behavior authenticity determination data stream, ending posture feature data, and posture reconstruction duration data are aligned along the time axis. For instantaneously generated static features (such as ending posture and reconstruction duration), the system uses zero-order preservation or temporal diffusion algorithms to broadcast their feature information within the corresponding action interval, fusing them into multi-channel input features containing logical, spatial, and temporal dimensions.

[0122] Multi-channel input features are fed into the behavior phase determination model to obtain the attack / defense phase classification results for each time step. Internally, the behavior phase determination model uses a temporal convolutional network (TCN) or a recurrent neural network with an attention mechanism to non-linearly extract the dynamic evolution of input features within a sliding window. The model not only focuses on the feature values ​​at the current time step but also incorporates the previously identified "authenticity" attribute to adjust the classification probability for subsequent phases. For example, if the authenticity is determined to be a "fake move," the model automatically redirects the subsequent phase recognition weights from the "closing phase" to the "transition phase." The model's output layer calculates the probability distribution for each input frame belonging to a preset phase such as "initiation phase, execution phase, impact phase, closing phase, and defense rebuilding phase."

[0123] The system extracts the classification results at each moment based on the principle of maximizing probability and arranges them in streaming order according to the original time sequence. To eliminate subtle jumps in the recognition results between frames, the system also applies mean-mode filtering or conditional random fields for temporal smoothing to ensure that the physical logic of phase switching conforms to common sense of human kinematics (e.g., the finishing move phase must be after the impact phase). The final data stream of attack and defense behavior phases is output in the form of a structured time-series report, clearly marking the start and end frame indices, duration, and corresponding feature confidence scores for each phase.

[0124] By fusing multi-channel features and employing a temporal classification model, this application achieves in-depth analysis of the entire lifecycle of Sanda (Chinese kickboxing) movements. By linking authenticity assessment, static posture, and temporal efficiency, the output stage data stream is no longer isolated labels but rather a technical evaluation report with causal logic. This output mechanism, based on a closed-loop multi-source data system, can accurately capture rapidly changing offensive and defensive transition points, thus providing trainees with comprehensive quantitative feedback from movement intention to execution efficiency, thereby enhancing the scientific level of Sanda teaching and tactical analysis.

[0125] The following is a specific implementation of a depth vision-based martial arts Sanda movement recognition training method:

[0126] A Sanda (Chinese kickboxing) athlete, equipped with a 3D visual capture system consisting of 8 high-speed depth cameras (sampling rate 120Hz), conducted combination sparring training in a standard ring environment, focusing on straight punches followed by roundhouse kicks. The system used the MediaPipe human posture algorithm to extract the 3D coordinate sequences of 25 key skeletal points throughout the body in real time. During the initial phase of a straight punch, the system captured a sagittal projection change rate of 18° / frame at the torso center point, and a peak torso acceleration vector magnitude of 7.2 m / s². 2After weighted fusion (each weight coefficient is 0.5), the fusion value is (7.2 + 18 × 0.1) / 2 = 4.5, which exceeds the preset fusion threshold Y = 3.8, generating a startup trend marker. For the fist end joint, the average cosine similarity of the instantaneous velocity vector calculated within a 10-frame sliding window is 0.91, higher than the preset consistency threshold of 0.88, and the spatial main direction displacement reaches 22cm, exceeding the preset displacement threshold. The association analysis module determined that the behavior was valid and initiated.

[0127] During the execution of the straight punch, the system detected a change rate of -2.1 m / s in the fist's trajectory speed. 2 The velocity vector dot product between adjacent frames abruptly changes from +0.87 to -0.62, indicating a reversal of motion direction. However, because the rate of velocity change is not lower than the preset threshold, the motion direction is not reversed. No action interruption flag was triggered. After the straight punch reaches its peak momentum, the system locates the attack end frame using the kinetic energy decay curve and extracts the supporting surface area of ​​0.08m² under that frame. 2 The distance between the center of gravity projection point and the center of the support surface is 0.12m, and the calculated stability geometric feature value S = 0.08 / 0.12 ≈ 0.67; the lateral deviation of the head from the central axis of the torso is 6cm, which is within the safe threshold range; the height of the body's center of gravity is 1.12m. These feature values ​​are normalized to generate the final posture feature data. The posture reconstruction module compares the standard defensive stance template frame by frame, starting from the attack end frame. When the similarity score is higher than 0.92 for four consecutive frames, the posture is determined to revert to a defensive stance, and the posture reconstruction time is calculated to be 0.82 seconds.

[0128] The system achieves time synchronization of multi-view skeletal data (error <2ms) through the NTP network time protocol. It applies Dynamic Time Warping (DTW) to align the initiation trend data stream (120Hz), trajectory continuity data stream (120Hz), and pose reconstruction duration data (single-frame scalar). Based on a behavior phase determination model pre-trained with a Bidirectional Long Short-Term Memory (Bi-LSTM) network, it inputs multi-channel fused features and outputs the corresponding temporal sequence of offensive and defensive behavior phase data streams. In the straight punch followed by a whip kick combination, the system accurately divides the phases into initiation (0-0.25s), execution (0.25-0.68s), recovery (0.68-0.95s), defense reconstruction (0.95-1.77s), and whip kick initiation (1.77-2.10s). Statistical analysis of the behavior authenticity determination results shows a 94.2% accuracy rate for identifying fake actions and a 96.8% recall rate for identifying real attacks. The comprehensive attack and defense effectiveness index is calculated using a dynamic weighting formula: E = 0.4 × (0.67 / 0.85) + 0.3 × (1 - 6 / 15) + 0.2 × (1 - 0.82 / 1.2) + 0.1 × 0.968 = 0.655. (Where, stability feature weight α = 0.4, head offset weight β = 0.3, reconstruction time weight γ = 0.2, and true attack weight δ = 0.1; 0.85 is the excellent stability feature value, 15cm is the safe threshold for head offset, and 1.2s is the excellent reconstruction time.)

[0129] The preset target value for the combined movement efficiency index of professional athletes is 0.72. The current value of 0.655 triggers a level 1 optimization prompt (Δ=0.065). The system generates targeted training adjustment suggestions: 1. Straight punch recovery phase center of gravity control training: increase single-leg support balance exercises to improve support surface stability; 2. Defensive posture transition training: introduce mirror simulation confrontation to shorten posture reconstruction time; 3. Feint intention reinforcement training: optimize trajectory continuity characteristics through alternating practice of feints and real movements.

[0130] Based on the behavioral phase prediction model of Temporal Convolutional Network (TCN), the phase division of the subsequent 5 sets of combined actions was simulated. The prediction shows that after targeted training, the posture reconstruction time can be shortened to 0.65s, the stability feature value can be improved to 0.78, and the comprehensive efficiency index is expected to reach 0.71, which is close to the level of professional athletes.

[0131] This embodiment demonstrates the complete application process of a depth vision-based martial arts Sanda movement recognition training system in actual combat. Through multi-dimensional motion feature extraction, intelligent intent judgment, and dynamic stage division, it realizes automated and refined analysis of Sanda offensive and defensive behaviors, providing objective and accurate data support for training effect evaluation and technical movement optimization.

[0132] Example 2:

[0133] like Figure 3 The diagram shown is a system connection diagram of this application. The martial arts Sanda movement recognition training system based on depth vision, using the aforementioned martial arts Sanda movement recognition training method based on depth vision, includes:

[0134] The data acquisition module is used to acquire the three-dimensional skeletal coordinate data sequence of the trainee in a continuous time sequence;

[0135] The data processing module is used to calculate and generate a starting trend data stream describing the overall movement trend of the body based on the three-dimensional skeletal coordinate data sequence, and to calculate and generate a trajectory continuity data stream describing the consistency of movement at the limb ends.

[0136] The correlation analysis module is used to perform correlation analysis on the initiation trend data stream and the trajectory continuity data stream to generate a behavior initiation judgment data stream;

[0137] The dynamic extraction module is used to jointly detect the direction and velocity of the trajectory continuity data stream and generate an action interruption data stream;

[0138] The intent determination module is used to perform logical operations based on the behavior initiation determination data stream and the action interruption data stream to generate a behavior authenticity determination data stream.

[0139] The attack and defense assessment module is used to locate the attack end frame in the three-dimensional skeleton coordinate data sequence based on the judgment results representing the real attack in the behavior authenticity judgment data stream, and extract the posture structure features of the attack end frame to generate end posture feature data.

[0140] The attitude reconstruction module is used to calculate the time interval from the end of the attack frame to reaching the preset defensive attitude and generate attitude reconstruction time data.

[0141] The phase fusion module is used to input the behavior authenticity judgment data stream, the end posture feature data, and the posture reconstruction duration data into the pre-trained behavior phase judgment model, and output the attack and defense behavior phase data stream corresponding to the time sequence.

[0142] This application constructs a closed-loop digital training and monitoring system through deep collaboration between the hardware perception layer, the data computing layer, and the logical decision-making layer.

[0143] At the hardware architecture level, the data acquisition module not only includes front-end 3D vision sensors (such as depth cameras with infrared active ranging capabilities or multi-view high-frame-rate industrial cameras), but also integrates a high-performance edge computing gateway. This gateway is equipped with a SoC chip with parallel computing power (such as an embedded GPU or NPU) for real-time preprocessing of the raw photoelectric signals. The sensor cluster ensures strict alignment of multi-view spatial coordinates through synchronous triggers, providing a low-latency, high-precision 3D skeletal coordinate reference for subsequent data processing modules. In addition, the system includes a storage unit and a high-speed bus interface for temporarily storing skeletal coordinate data sequences, ensuring the data retrieval bandwidth for the attack and defense assessment module when tracing back to the end frame of an attack.

[0144] As a typical application example, when a trainee performs a combination of "lead hand jab followed by rear hand straight punch" within the sensor's field of view: the data acquisition module captures the coordinate flow of key points throughout the body; the data processing module calculates the center of gravity acceleration at the initiation of the lead hand jab. The correlation analysis module identifies the lead hand initiation trend marker, but the dynamic extraction module subsequently detects that the lead hand trajectory reverses direction and experiences a sudden drop in kinetic energy before reaching effective attack depth, generating an action interruption marker. Based on this logic, the intent determination module calculates that the jab is a "feint" and transmits the authenticity marker to the stage fusion module, avoiding deep power consumption calculations for invalid induced actions. When the rear hand straight punch is actually delivered, the intent determination module determines it as a "real attack." The attack and defense evaluation module immediately locks the attack end frame at the downward point after the attack reaches its momentum peak and extracts the trainee's body support area and center of gravity height at that moment. The posture reconstruction module simultaneously starts timing until it detects the trainee returning to the standard defensive stance template. Finally, the phase fusion module integrates the authenticity of the two actions, the quality of the finishing posture of the rear hand punch, and the duration of the defensive return to position, and generates a complete offensive and defensive timing report in real time on the display terminal.

[0145] This system combines high-frequency sampling at the hardware level with a logical discrimination matrix at the software level to automate the decomposition of complex game-playing behaviors in Sanda (Chinese kickboxing). The serial triggering mechanism between system modules reduces the processor's computational load on unnecessary data, while the three-dimensional evaluation of "intention-posture-duration" gives the system's output behavioral phase data stream extremely high practical reference value. This not only improves the robustness of action recognition but also provides a quantitative closed-loop feedback tool for optimizing the rhythm of athletes' offensive and defensive transitions.

[0146] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1. A martial arts Sanda (Chinese kickboxing) movement recognition training method based on depth vision, characterized in that, Includes the following steps: Acquire the 3D skeletal coordinate data sequence of the trainee within a continuous temporal sequence; Based on the three-dimensional skeletal coordinate data sequence, a starting trend data stream describing the overall movement trend of the body is calculated and generated, and a trajectory continuity data stream describing the consistency of movement at the limb ends is calculated and generated. A correlation analysis is performed between the initiation trend data stream and the trajectory continuity data stream to generate a behavior initiation determination data stream; The trajectory continuity data stream is subjected to joint detection of direction and velocity to generate an action interruption data stream; Logical operations are performed on the behavior initiation determination data stream and the action interruption data stream to generate a behavior authenticity determination data stream; Based on the determination results representing real attacks in the behavior authenticity determination data stream, the attack end frame is located in the three-dimensional skeleton coordinate data sequence, and the posture structure features of the attack end frame are extracted to generate end posture feature data. Calculate the time interval from the end of the attack frame to reaching the preset defensive posture, and generate posture reconstruction time data; The behavior authenticity determination data stream, the ending posture feature data, and the posture reconstruction duration data are input into the pre-trained behavior stage determination model, and the attack and defense behavior stage data stream corresponding to the time sequence is output.

2. The martial arts Sanda movement recognition training method based on depth vision according to claim 1, characterized in that, The specific steps involved in calculating and generating the initiation trend data stream describing the overall movement trend of the body include: For the coordinate sequence of the torso center point in the three-dimensional skeleton coordinate data sequence, calculate the displacement vector between consecutive frames; Perform a difference operation on the displacement vector to obtain an acceleration vector sequence; Calculate the rate of change of the projection of the trunk center point relative to its position at the previous moment onto the sagittal plane of the human body, and use it as the equivalent rate of change of posture of the trunk around the supporting joint. The magnitude of the acceleration vector sequence is weighted and fused with the equivalent attitude change rate to generate a fusion value. When the fusion value exceeds a preset fusion threshold, a start-up trend marker is generated.

3. The martial arts Sanda movement recognition training method based on depth vision according to claim 1, characterized in that, The specific steps for calculating and generating a trajectory continuity data stream describing the consistency of movement at the limb extremities include: Calculate the instantaneous velocity vector for the coordinate sequence of the distal joints in the three-dimensional skeletal coordinate data sequence; Within a sliding window of a preset length of M frames, the average cosine similarity between the instantaneous velocity vector in the current frame and the historical frames within the window is calculated as the directional consistency coefficient. The displacement sequence of the limb distal joints is subjected to high-pass filtering to filter out motion components below a preset frequency threshold and generate smoothed trajectory data. Based on the directional consistency coefficient and the smoothed trajectory data, a trajectory continuity data stream is generated.

4. The martial arts Sanda movement recognition training method based on depth vision according to claim 3, characterized in that, The specific steps for performing correlation analysis between the initiation trend data stream and the trajectory continuity data stream to generate the behavior initiation determination data stream include: Detect the startup trend marker in the startup trend data stream; Starting from the moment when the start trend marker is detected, the trajectory continuity data stream is acquired within a preset start association window; Determine whether the directional consistency coefficient in the trajectory continuity data stream is higher than a preset consistency threshold within the preset startup association window, and whether the displacement of the smoothed trajectory data in the main spatial direction exceeds a preset displacement threshold. If both conditions are met, a behavior initiation determination flag is generated and used as the behavior initiation determination data stream.

5. The martial arts Sanda movement recognition training method based on depth vision according to claim 4, characterized in that, The specific steps for jointly detecting direction and velocity in the trajectory continuity data stream to generate an action interruption data stream include: Calculate the rate of change of the smoothed trajectory data between consecutive frames; The smoothed trajectory data is detected to have a reversed motion direction between consecutive frames; wherein, the reversal of motion direction is determined by the dot product of the velocity vectors of adjacent frames changing from positive to negative; When the rate of change of speed is detected to be lower than a preset threshold and the direction of motion is simultaneously reversed, an action interruption flag is generated. Based on the steepness of the decrease in the rate of change of velocity and the range of the angle of reversal of the direction of motion, a type label is assigned to the action interruption marker, forming the action interruption data stream.

6. The martial arts Sanda movement recognition training method based on depth vision according to claim 4, characterized in that, The specific steps for generating a behavior authenticity determination data stream by performing logical operations based on the behavior initiation determination data stream and the action interruption data stream include: Align the behavior initiation determination flag in the behavior initiation determination data stream with the action interruption flag in the action interruption data stream in terms of timing; If at least one action interruption marker is detected after the action initiation marker within the preset authenticity determination window, the action initiation is determined to be a fake action; if no action interruption marker is detected, it is determined to be a real attack. The generated fake actions and the real attack behaviors are arranged in chronological order according to the corresponding behavior initiation judgment markers to form the behavior authenticity judgment data stream.

7. The martial arts Sanda movement recognition training method based on depth vision according to claim 1, characterized in that, The specific steps for extracting the pose structure features of the attack termination frame and generating termination pose feature data include: Based on the three-dimensional skeletal coordinate data sequence of the attack end frame, calculate at least one geometric feature value characterizing body stability, at least one distance feature value characterizing head offset, and body center of gravity height value. The geometric feature value, the distance feature value, and the center of gravity height value are combined to generate the final posture feature data.

8. The martial arts Sanda movement recognition training method based on depth vision according to claim 1, characterized in that, The specific steps for calculating the time interval from the end of the attack frame to reaching the preset defensive posture, and generating posture reconstruction time data, include: A predefined standard defensive stance template; the standard stance template is composed of the joint coordinates representing the standard defensive stance. Starting from the frame where the attack ends, calculate the similarity score between the bone coordinates of the current frame and the standard defensive posture template frame by frame. When the similarity score is higher than the preset similarity threshold for N consecutive frames, it is determined that the preset defensive posture has been reached; where N is an integer greater than 1. The time difference between the attack end frame and the first frame where the target is met is used as the attitude reconstruction duration data.

9. The martial arts Sanda movement recognition training method based on depth vision according to claim 1, characterized in that, The specific steps for outputting the attack and defense behavior phase data stream corresponding to the time series include: The behavior authenticity determination data stream, the ending posture feature data, and the posture reconstruction duration data are aligned along the time axis and fused into a multi-channel input feature. The multi-channel input features are input into the behavior stage determination model to obtain the attack and defense stage classification results of the input features at each time step. The attack and defense phase classification results are arranged in chronological order to form the attack and defense behavior phase data stream.

10. A martial arts Sanda (Chinese kickboxing) movement recognition training system based on depth vision, characterized in that, The method for training martial arts Sanda movement recognition based on depth vision as described in any one of claims 1 to 9 includes: The data acquisition module is used to acquire the three-dimensional skeletal coordinate data sequence of the trainee in a continuous time sequence; The data processing module is used to calculate and generate a starting trend data stream describing the overall movement trend of the body based on the three-dimensional skeletal coordinate data sequence, and to calculate and generate a trajectory continuity data stream describing the consistency of movement at the limb ends. The correlation analysis module is used to perform correlation analysis on the initiation trend data stream and the trajectory continuity data stream to generate a behavior initiation judgment data stream; The dynamic extraction module is used to perform joint detection of direction and velocity on the trajectory continuity data stream to generate an action interruption data stream; The intent determination module is used to perform logical operations on the behavior start determination data stream and the action interruption data stream to generate a behavior authenticity determination data stream. The attack and defense assessment module is used to locate the attack end frame in the three-dimensional skeleton coordinate data sequence based on the judgment result representing the real attack in the behavior authenticity judgment data stream, and extract the posture structure features of the attack end frame to generate end posture feature data. The attitude reconstruction module is used to calculate the time interval from the end of the attack frame to the achievement of the preset defensive attitude, and generate attitude reconstruction time data. The phase fusion module is used to input the behavior authenticity determination data stream, the ending posture feature data, and the posture reconstruction duration data into the pre-trained behavior phase determination model, and output the attack and defense behavior phase data stream corresponding to the time sequence.

Images