A martial arts action recognition method based on motion capture

Abstract

The application relates to the field of motion capture technology, and particularly discloses a martial art motion recognition method based on motion capture, which comprises the following steps: acquiring original motion capture data of a user at a current sampling moment; generating a reference plane based on plantar pressure distribution data; classifying pixel points in a depth image according to the spatial distance between the pixel points and the reference plane, extracting a pixel point set D belonging to the user's body entity after classification, calculating the vertical projection contour of the pixel point set D on the reference plane, and marking the vertical projection contour as a real limb contour; generating a posture snapshot at the current moment based on the real limb contour and joint coordinate data; comparing the posture snapshot with a pre-stored standard martial art motion template library, and outputting the martial art motion name at the current sampling moment and the deviation between the posture snapshot and the standard template. The application improves the motion recognition accuracy in a home exercise scene.

Description

Technical Field

[0001] This invention relates to the field of motion capture technology, and more specifically to a method for recognizing martial arts movements based on motion capture. Background Technology

[0002] In home workout scenarios, the practical difficulties faced by motion recognition algorithms generally stem from the disconnect between their pre-designed ideal conditions and the user's real environment. Algorithms typically assume rigid support surfaces and tight clothing as a basis to ensure the stability of the reference coordinate system and the accurate extraction of limb contours.

[0003] However, users often practice on non-rigid surfaces such as carpets or mattresses. These support surfaces undergo elastic deformation under body pressure, dynamically changing the body's center of gravity and causing continuous deviations in the algorithm's calculations based on a fixed rigid benchmark. Simultaneously, the material of loose clothing causes natural drooping and swaying, forming virtual silhouettes that do not match real limbs. The trajectories extracted by the visual algorithm through edge detection are actually magnified and delayed motion patterns of the clothing.

[0004] In this scenario, deformation of the support surface first causes coordinate system shift, misleading the initial positioning of the limbs; interference from clothing outlines further distorts the extraction of the motion trajectory, creating superposition errors. Ultimately, the algorithm's recognition results deviate significantly from the user's actual actions, and even standard actions may be misjudged as non-standard, severely weakening the accuracy and reliability of action recognition. Summary of the Invention

[0005] The purpose of this invention is to provide a martial arts movement recognition method based on motion capture, thereby solving the aforementioned technical problems.

[0006] The objective of this invention can be achieved through the following technical solutions:

[0007] A method for martial arts movement recognition based on motion capture includes the following steps:

[0008] Acquire the user's raw motion capture data at the current sampling time. The raw motion capture data includes depth image, joint coordinate data, and plantar pressure distribution data.

[0009] A reference plane is generated based on plantar pressure distribution data;

[0010] The pixels are classified according to the spatial distance between the pixels in the depth image and the reference plane. The set D of pixels belonging to the user's body entity after classification is extracted. The vertical projection contour of the pixel set D on the reference plane is calculated and the vertical projection contour is recorded as the real limb contour.

[0011] Generate a snapshot of the current pose based on real limb contours and joint coordinate data;

[0012] The posture snapshot is compared with a pre-stored library of standard martial arts movement templates, and the name of the martial arts movement at the current sampling moment and the deviation between the posture snapshot and the standard template are output.

[0013] As a further aspect of the present invention: generating the reference plane includes:

[0014] Extract the spatial coordinates and sensed pressure values ​​of all pressure sensing units from the plantar pressure distribution data;

[0015] The spatial coordinates of the pressure sensing unit are weighted and averaged using the sensed pressure value as the weight to obtain the three-dimensional coordinates of the pressure center point at the current sampling time.

[0016] Based on the spatial coordinates of the pressure sensing unit and the sensed pressure value, the principal axis direction of the pressure distribution is calculated.

[0017] A reference plane is established, with the normal vector of the reference plane pointing in the direction of the principal axis, and the reference plane includes the three-dimensional coordinates of the pressure center point.

[0018] As a further aspect of the present invention: calculating the vertical projection profile includes:

[0019] Map each pixel in the depth image to a 3D space to obtain a spatial point, and calculate the directed distance of each spatial point to the reference plane;

[0020] All spatial points whose absolute value of directed distance falls within a first set range are included in the candidate body entity point set;

[0021] Clustering is performed on all spatial points in the candidate body entity point set to obtain several clusters. The clustering standard is Euclidean distance, and the cluster with the largest number of spatial points is taken as the pixel point set D.

[0022] The outer boundary of the closed region formed by drawing perpendicular lines from spatial points in the pixel set D to the reference plane, with the feet of all perpendiculars on the reference plane, is the vertical projection profile.

[0023] As a further aspect of the present invention: extracting a pose snapshot includes:

[0024] Extracting pose snapshots includes:

[0025] Extract the geometric center point of the real limb contour;

[0026] Establish a three-dimensional skeletal model containing all joints, wherein the skeletal joint coordinate data includes the initial coordinates of each joint in the three-dimensional skeletal model;

[0027] An objective function for optimization is established based on the geometric center point and the 3D skeletal model. The minimum value of the objective function is solved, and the optimal coordinates of all relevant nodes in the 3D skeletal model are calculated. The optimal coordinates of all joints are used as a pose snapshot at the current sampling time.

[0028] As a further aspect of the present invention: the deviation between the martial arts action name and posture snapshot at the current sampling moment and the standard template includes:

[0029] Calculate the positional difference between the optimal coordinates of each joint point in the posture snapshot and the coordinates of the corresponding joint point in each standard posture template in the standard martial arts movement template library;

[0030] The positional differences are summed to obtain the total difference between the pose snapshot and each standard pose template.

[0031] Find the standard posture template D that minimizes the sum of differences;

[0032] The name of the standard posture template D is used as the name of the martial arts movement at the current sampling moment;

[0033] The sum of these differences is recorded as the deviation between the pose snapshot and the standard template.

[0034] As a further aspect of the present invention, it also includes:

[0035] The deviations between the posture snapshots corresponding to the same martial arts move name and the standard template are sorted in chronological order to obtain a deviation sequence, and the deviation sequence is then visualized.

[0036] As a further aspect of the present invention: if there are two or more clusters with the same number of spatial points and the maximum value, then the original motion capture data at the current sampling time will not be processed in subsequent steps.

[0037] The beneficial effects of this invention compared to the prior art are as follows:

[0038] This invention generates a reference plane based on plantar pressure distribution that updates synchronously with changes in the user's force state. This eliminates the reliance on fixed support assumptions during action recognition, maintaining a stable spatial reference for posture analysis even when deformation occurs on non-rigid surfaces. Through joint processing of depth images and the reference plane, only spatial point groups corresponding to the human body are retained, forming a realistic limb contour. This prevents false shapes caused by clothing movement from being included in subsequent posture calculations. A 3D skeletal model is constructed using the realistic limb contour and joint coordinate data. Optimal coordinates for each joint are obtained through optimization, ensuring the posture snapshot closely matches the actual movement structure of the human body, thus reducing overall posture errors caused by initial joint point offsets. Furthermore, the posture snapshot is compared with a standard action template to obtain the action name and posture deviation, allowing the recognition results to more accurately reflect the user's actual movement. Through this processing, this invention can achieve relatively stable human posture and action category judgments even in complex home environments, mitigating the superposition errors caused by support surface deformation and clothing interference, and improving the reliability and consistency of action recognition results. Attached Figure Description

[0039] The invention will now be further described with reference to the accompanying drawings.

[0040] Figure 1 This is a flowchart illustrating a martial arts movement recognition method based on motion capture according to the present invention. Detailed Implementation

[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] Please see Figure 1 As shown, this invention is a martial arts movement recognition method based on motion capture, comprising the following steps:

[0043] Acquire the user's raw motion capture data at the current sampling time. The raw motion capture data includes depth image, joint coordinate data, and plantar pressure distribution data.

[0044] A reference plane is generated based on plantar pressure distribution data;

[0045] In a preferred embodiment of the present invention, generating a reference plane includes:

[0046] The installation position of each pressure unit in the pressure sensing pad in the three-dimensional coordinate system is used as the fixed calibration coordinate, and the pressure value corresponding to the sampling time is used as the force information of the unit. Then, a set of spatial position and pressure value correspondence is formed by reading them one by one.

[0047] After obtaining this data, the spatial coordinates of the pressure sensing units are weighted and averaged using the sensed pressure value to obtain the three-dimensional coordinates of the pressure center point. This links the spatial position of each sensing unit to its pressure value, making areas with higher pressure have a stronger influence on the overall position estimation. For example, when a user applies significant pressure to the forefoot area of ​​one foot while standing, the spatial coordinates of that area will contribute more to the position of the pressure center point during the weighting process, thus making the pressure center point closer to the user's actual force application location.

[0048] After determining the location of the pressure center point, to obtain the principal axis direction of the pressure distribution, it is necessary to determine the main extension direction of the pressure distribution using the spatial coordinates and pressure values ​​of each pressure sensing unit. Principal component analysis can be used, with the pressure value as the weight source in this method, making the influence of the high-stress region on the spatial direction more significant during the calculation. This method can determine the direction with the largest variance in the pressure distribution, which is then taken as the principal axis direction of the pressure distribution.

[0049] After obtaining the pressure center point and the principal axis direction of the pressure distribution, the construction of the reference plane is based on these two quantities. The normal vector of the reference plane is consistent with the principal axis direction of the pressure distribution, so that the spatial orientation of the plane can be associated with the force structure of the user's foot. At the same time, the reference plane needs to contain the three-dimensional coordinates of the pressure center point. By taking the pressure center point as a known point on the plane, the reference plane has a definite position in space.

[0050] The reference plane is no longer a fixed, horizontal, abstract "floor," but a dynamic plane bound to the current stress state. Even if the user sinks or tilts on the soft support surface, the position and orientation of the plane will be adjusted in real time according to the pressure distribution, so that all subsequent projections and distance calculations related to the plane are always based on the actual stress posture, avoiding overall coordinate drift caused by non-rigid support surfaces.

[0051] The pixels are classified according to the spatial distance between the pixels in the depth image and the reference plane. The set D of pixels belonging to the user's body entity after classification is extracted. The vertical projection contour of the pixel set D on the reference plane is calculated and the vertical projection contour is recorded as the real limb contour.

[0052] In another preferred embodiment of the present invention, calculating the vertical projection profile includes:

[0053] By using the known imaging parameters of the depth camera, the row and column coordinates of pixels in the image plane are converted into the horizontal and vertical positions in the camera coordinate system. The distance from the point to the camera is determined by using the depth value of the corresponding pixel in the depth image, so that each pixel can obtain a definite position in three-dimensional space.

[0054] After mapping, to remove outliers that are clearly unrelated to the human body region, the directed distance from each spatial point to the reference plane is calculated, allowing the relative position of each point in the normal direction of the reference plane to be quantified. The directed distance represents the distribution of spatial points in the vertical direction of the reference plane. Points with an absolute value of directed distance within a first set range are assigned to the candidate body entity point set according to pre-defined parameters, excluding extreme points that are too large or too small from the reference plane. For example, discrete points caused by depth noise or background points in the environment that are far from the human body are usually not in this range and cannot enter the candidate set.

[0055] After directed distance filtering, the candidate point set may still contain some clothing movement points or local environmental points adjacent to the human body shape. Although these points are close to parts of the body in height, they often lack continuity in spatial structure. To further separate the point cloud corresponding to the human body entity from the candidate set, clustering is performed on the candidate points based on Euclidean distance. By judging the spatial distance relationship between the points, continuously distributed points are divided into the same cluster. Points on the human body surface usually show a large continuous distribution in three-dimensional space, while points formed by clothing movement are usually local, scattered, and not patchy. After clustering, several clusters can be obtained from the candidate set. The cluster with the most points is selected as the pixel set D based on the characteristic that the human body surface forms the main region in the depth image, so that this set represents the three-dimensional point cloud of the user's actual body surface.

[0056] After obtaining the pixel set D, to transform the three-dimensional spatial information into a two-dimensional contour for easier overall analysis, a perpendicular line is drawn from each spatial point in set D along the normal direction of the reference plane, so that each spatial point corresponds to a perpendicular foot point on the reference plane. By aggregating all the perpendicular foot points to form a point set on the reference plane, the outer boundary of the closed region is determined within the planar distribution of this point set. The determination of the outer boundary can be achieved by performing a contour tracing operation on the point set on the reference plane, so that all the perpendicular foot points scattered on the plane form an envelope region according to their spatial connections. The outermost trajectory of this envelope region is the vertical projection contour, which reflects the planar projection shape of the user's body shape on the reference plane at the current sampling moment.

[0057] Understandably, if there are two or more clusters with the same number of spatial points and the maximum number of points, the raw motion capture data at the current sampling time will not be processed in subsequent steps.

[0058] Spatial consistency alignment is performed based on real limb contours and joint coordinate data, and a pose snapshot at the current sampling time is extracted after alignment.

[0059] In another preferred embodiment of the present invention, extracting a pose snapshot includes:

[0060] Geometric center points are extracted from the actual limb contours to obtain reference information that reflects the user's overall spatial position at the current sampling time. These geometric center points serve as the basis for shape constraints in subsequent pose solving.

[0061] After acquiring joint coordinate data, a 3D skeletal model is constructed for the current user. Each joint coordinate from the data is assigned to its corresponding joint in the 3D skeletal model, ensuring the model contains the initial spatial positions of each joint in the user's current posture from the outset. This results in a 3D skeletal model with actual human structural topology and specific initial joint coordinates. During subsequent optimization, positional constraints are established between this model and the actual limb contour, allowing the initial joint coordinates to be adjusted based on shape information during the optimization process, ultimately leading to final joint positions that better reflect the user's actual posture.

[0062] After the 3D skeletal model is established, an optimization objective function is constructed, constrained by the positional relationship between the geometric center point and the 3D skeletal model. This objective function constrains the spatial relationship between the projected positions of joints on the reference plane and the actual limb contour, while simultaneously constraining the relative positional relationship between adjacent joints. This ensures that the 3D skeletal model both approximates the actual contour and maintains a reasonable human skeletal structure during optimization. For example, if a joint's initial coordinates deviate from its actual position due to occlusion, the optimization objective function is solved iteratively. In each iteration, the coordinates of that joint and other joints are updated, gradually adjusting the positions of all joints to a state more consistent with the actual limb contour and human structure. When the optimization objective function converges to its minimum, the optimal coordinates of all joints in the 3D skeletal model are obtained. This set of optimal joint coordinates is used as the pose snapshot at the current sampling time.

[0063] Understandably, the vertical projection profile originates from the outer boundary of the projected 3D point cloud onto the reference plane after filtering and clustering. This outer boundary reflects the maximum planar area occupied by the human body at that moment, describing the macroscopic geometric envelope structure of the body shape. The 3D skeletal model consists of the 3D coordinates of each joint. These coordinates need to be constrained by the shape of the human body during pose solving, ensuring that the joints do not fall into positions inconsistent with the real body in space. Although the projection profile does not provide information about the internal skeletal structure, it provides the skeletal model with a geometric constraint space consistent with the real human body shape. This allows the skeletal model to adjust itself during optimization based on the relationship between the projected positions of the joints on the reference plane and the projection profile, preventing the skeletal structure from exceeding the external contour of the human body or failing to find a reasonable pose within the scope defined by the contour. In this way, the projection profile provides the overall boundary of the body shape, and the skeletal model provides the joint structure and initial pose. Constraints are established between the two through an optimization objective function, ensuring consistency between the skeletal pose and the projection profile, thereby achieving correction and supplementation of the skeletal pose based on shape information.

[0064] Specifically, the objective function for optimization is as follows:

[0065]

[0066] Where P represents a preset set of key points, X i V represents the coordinates to be optimized for the i-th joint in the set of joints P. k V represents the coordinates to be optimized for the k-th joint in the 3D skeleton model. torso This represents the initial coordinates of preset joints in the 3D skeleton model. X represents i The coordinates mapped to the reference plane, where B represents the actual limb contour. express The shortest distance to the actual limb outline B (denoted as the first distance). This represents the summation of the squares of the first distances corresponding to all joints in the joint set P, where M represents the total number of joints in the 3D skeleton model, λ1 and λ2 are preset correction coefficients, and Y... k This represents the initial coordinates of the k-th joint in the 3D skeleton model (not the optimal or unoptimized coordinates; the function contains only two unoptimized variables, X). i With V k ), c represents the coordinates of the geometric center point;

[0067] Understandably, the coordinates to be optimized gradually approach the most reasonable position that simultaneously satisfies all constraints through iterative updates during the solution of the objective function. The objective function consists of multiple error terms reflecting the geometric relationships of the human body. The coordinates to be optimized are initially taken from the initial joint positions of the skeletal model. As the solution process unfolds, each iteration adjusts these coordinates according to the magnitude of the current error, moving them towards the direction with the smaller error. Error terms related to the real limb contour cause the projection of the joint points on the reference plane to approach the outer boundary of the real contour; error terms related to the consistency of the skeletal structure cause adjacent joint points to maintain a relative relationship consistent with the human body topology; and constraints related to the geometric center point ensure that the overall skeleton posture is consistent with the actual body position. As iterations continue, the error gradually decreases. When the value of the objective function reaches a state where it can no longer decrease significantly, the coordinates to be optimized converge to a stable solution under the combined effect of all constraints. The coordinates to be optimized at this moment are the optimal coordinates, used to represent the joint positions that best conform to the real body posture at that sampling time.

[0068] Note that X i Once the position in three-dimensional space is determined, the reference plane is defined by its normal direction and the pressure center point located on the plane. During the mapping process, X... i A straight line extends along the normal direction of the reference plane, intersecting the reference plane defined by the normal direction and the center of pressure. The point of intersection is X. i Corresponding position on the reference plane Since the position of the reference plane depends on the pressure center point, its spatial position is fixed. Therefore, the intersection point of Xi and the reference plane is unique, which can provide a clear and stable planar projection point for subsequent distance calculation.

[0069] It should be noted that the objective function is designed to find the set of coordinates among numerous possible joint coordinates that best approximates the actual human posture, allowing the skeletal model to naturally conform to the spatial shape of the real limb contour. This objective function consists of three parts, each bearing different constraints. These three parts work together to ensure that the solution process simultaneously satisfies three requirements: conforming to the contour, maintaining a reasonable skeletal structure, and maintaining trunk stability. The first condition ensures a realistic fit, the second ensures the skeletal structure is not damaged, and the third ensures overall center stability. The combined effect of these three factors ensures that the solved joint coordinates conform to human structural relationships, closely approximate the actual movement contour, and avoid overall displacement caused by interference from the supporting surface and clothing. Because these constraints coexist, the joint position corresponding to the minimum value of the objective function can be considered a reasonable estimate of the human posture at the current moment. Specifically:

[0070] The first term represents the degree of fit between the joints of the skeletal model projected onto the reference plane and the actual limb contour. The actual limb contour reflects the body's shape boundary on the plane during user movements. By mapping the joints of the skeletal model onto the reference plane, we can determine the deviation between the projected position of the joint and the actual limb shape. The greater the distance, the more significantly the skeletal model deviates from the actual human shape in that area. By minimizing this term, we are essentially pushing the overall shape of the skeletal model closer to the actual contour, so that the optimized posture no longer depends solely on the joints themselves, but rather on the shape information provided by the actual contour for correction. This reduces joint misalignment caused by interfering factors such as loose clothing, making the skeletal shape more consistent with actual human movements.

[0071] The second term expresses the spatial harmony of the skeletal structure itself. There are natural relationships between joints in the skeletal model; for example, the relative positions of adjacent joints should not exhibit sudden jumps or twists that contradict human anatomy. This term constrains the spatial relationships between joints within the skeletal model, preventing situations where local points satisfy contour fit but the overall skeleton deforms abnormally during the solution process. Optimization requires balancing both the first and second terms to ensure the skeleton closely resembles the realistic contour while maintaining an internal structure consistent with natural human posture. This term ensures that the final posture snapshot is not a set of isolated points, but rather a skeleton with consistent body structure.

[0072] The third term is used to maintain the proper position of the torso in the global coordinate system. The torso is the most stable and core part of human movement; if its position on the reference plane is significantly offset, the entire skeletal posture will be distorted. This term compares the relationship between the torso position in the skeletal model and the geometric center point of the contour, ensuring the stability of the central region of the human body during the solution process. This prevents the skeleton from being incorrectly pulled as a whole due to local contour changes. This avoids the overall drift effect caused by deformation of the support surface, making the final posture snapshot more accurately reflect the user's true body position.

[0073] The posture snapshot is compared with a pre-stored library of standard martial arts movement templates, and the name of the martial arts movement at the current sampling moment and the deviation between the posture snapshot and the standard template are output.

[0074] In a preferred embodiment of the present invention, obtaining the deviation between the martial arts movement name and posture snapshot at the current sampling moment and the standard template includes:

[0075] Calculate the positional difference between the optimal coordinates of each joint point in the posture snapshot and the coordinates of the corresponding joint point in each standard posture template in the standard martial arts movement template library;

[0076] The positional differences are summed to obtain the total difference between the pose snapshot and each standard pose template.

[0077] Find the standard posture template D that minimizes the sum of differences;

[0078] The name of the standard posture template D is used as the name of the martial arts movement at the current sampling moment;

[0079] The sum of these differences is recorded as the deviation between the pose snapshot and the standard template.

[0080] It should be noted that the deviations between the posture snapshots corresponding to the same martial arts movement name and the standard template are sorted in chronological order to obtain a deviation sequence, and the deviation sequence is then visualized.

[0081] This invention addresses the motion recognition deviations caused by support surface deformation and clothing interference during home exercise. By generating a reference plane that reflects the user's real-time force state based on the foot pressure distribution at each sampling moment, the spatial reference on which posture analysis depends is automatically adjusted with changes in the support surface, thus avoiding coordinate drift problems caused by fixed references from the source.

[0082] After acquiring the depth image, spatial points are filtered by a reference plane and grouped to determine body entity point groups, so that clothing swinging areas unrelated to the human body are not included in the contour extraction process, thereby reducing the impact of false shapes on the judgment of limb structure at the visual level.

[0083] Based on selected real limb contours, a pose snapshot conforming to human anatomy is obtained through optimized solution of skeletal models and joint coordinates. This corrects errors caused by clothing occlusion or environmental factors at the initial joint points, resulting in a stable posture expression consistent with actual movements. Action recognition ultimately relies on comparing the pose snapshot with a standard template. Therefore, the entire recognition process is built upon a dynamic reference plane, real limb contours, and optimized skeletal posture. This prevents deformation of the support surface and clothing interference from further accumulating errors in the recognition process, effectively solving the technical problem of decreased action recognition accuracy in home workout scenarios.

[0084] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the present invention should still fall within the scope of the present invention.

Claims

1. A martial arts movement recognition method based on motion capture, characterized in that, Includes the following steps: Acquire the user's raw motion capture data at the current sampling time. The raw motion capture data includes depth image, joint coordinate data, and plantar pressure distribution data. A reference plane is generated based on plantar pressure distribution data; The pixels are classified according to the spatial distance between the pixels in the depth image and the reference plane. The set D of pixels belonging to the user's body entity after classification is extracted. The vertical projection contour of the pixel set D on the reference plane is calculated and the vertical projection contour is recorded as the real limb contour. Generate a snapshot of the current pose based on real limb contours and joint coordinate data; The posture snapshot is compared with a pre-stored library of standard martial arts movement templates, and the name of the martial arts movement at the current sampling moment and the deviation between the posture snapshot and the standard template are output. Extracting pose snapshots includes: Extract the geometric center point of the real limb contour; Establish a 3D skeletal model containing all joints, and the skeleton joint coordinate data includes the initial coordinates of each joint in the 3D skeletal model; An objective function for optimization is established based on the geometric center point and the 3D skeletal model. The minimum value of the objective function is solved, and the optimal coordinates of all relevant nodes in the 3D skeletal model are calculated. The optimal coordinates of all joints are used as a pose snapshot at the current sampling time.

2. The martial arts movement recognition method based on motion capture according to claim 1, characterized in that, Generating the reference plane includes: Extract the spatial coordinates and sensed pressure values ​​of all pressure sensing units from the plantar pressure distribution data; The spatial coordinates of the pressure sensing unit are weighted and averaged using the sensed pressure value as the weight to obtain the three-dimensional coordinates of the pressure center point at the current sampling time. Based on the spatial coordinates of the pressure sensing unit and the sensed pressure value, the principal axis direction of the pressure distribution is calculated. A reference plane is established, with the normal vector of the reference plane pointing in the direction of the principal axis, and the reference plane includes the three-dimensional coordinates of the pressure center point.

3. The martial arts movement recognition method based on motion capture according to claim 1, characterized in that, Calculating the vertical projection profile includes: Map each pixel in the depth image to a 3D space to obtain a spatial point, and calculate the directed distance of each spatial point to the reference plane; All spatial points whose absolute value of directed distance falls within a first set range are included in the candidate body entity point set; Clustering is performed on all spatial points in the candidate body entity point set to obtain several clusters. The clustering standard is Euclidean distance, and the cluster with the largest number of spatial points is taken as the pixel point set D. The outer boundary of the closed region formed by drawing perpendicular lines from spatial points in the pixel set D to the reference plane, with the feet of all perpendiculars on the reference plane, is the vertical projection profile.

4. The martial arts movement recognition method based on motion capture according to claim 3, characterized in that, The deviation between the martial arts movement name and posture snapshot at the current sampling moment and the standard template includes: Calculate the positional difference between the optimal coordinates of each joint point in the posture snapshot and the coordinates of the corresponding joint point in each standard posture template in the standard martial arts movement template library; The positional differences are summed to obtain the total difference between the pose snapshot and each standard pose template. Find the standard posture template D that minimizes the sum of differences; The name of the standard posture template D is used as the name of the martial arts movement at the current sampling moment; The sum of these differences is recorded as the deviation between the pose snapshot and the standard template.

5. The martial arts movement recognition method based on motion capture according to claim 4, characterized in that, Also includes: The deviations between the posture snapshots corresponding to the same martial arts move name and the standard template are sorted in chronological order to obtain a deviation sequence, and the deviation sequence is then visualized.

6. The martial arts movement recognition method based on motion capture according to claim 3, characterized in that, If there are two or more clusters with the same number of spatial points and the maximum number of points, the raw motion capture data at the current sampling time will not be processed in subsequent steps.

Images