Gait recognition method based on Pluecker straight line
By modeling the motion features of bones and joints using dual quaternions and Plück coordinates, hierarchical gait features are constructed, solving the problems of incomplete gait recognition and insufficient feature discrimination in existing technologies, and achieving high-precision and robust identity recognition.
Patent Information
- Application Number
- CN202511793917.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-02
- Publication Date
- 2026-03-03
AI Technical Summary
Existing 3D gait recognition methods suffer from incomplete motion feature modeling and insufficient feature discrimination, making it impossible to effectively distinguish between overall motion and local relative motion. Furthermore, they are sensitive to individual differences, which affects recognition accuracy.
Dual quaternions and Plück coordinates are used to model the absolute motion of bones and the relative motion of joints, respectively, to construct hierarchical gait features, which are then identified using a deep learning model.
It achieves high-precision and robust gait recognition, overcomes the effects of perspective changes and body size differences, and improves the accuracy of identity recognition.
Smart Images
Figure CN121600596A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of computer vision and biometric recognition technology, and particularly to a gait recognition method based on Plück lines. Specifically, it is a method and system for gait feature extraction and recognition based on a three-dimensional human skeleton sequence, utilizing dual quaternions and Plück coordinates. More specifically, this invention provides a solution for achieving high-precision and robust identity recognition by separately modeling the absolute motion of bones and the relative motion of joints. Background Technology
[0002] Gait recognition, as a non-contact, long-range biometric identification technology, has unique application advantages. However, most existing 3D methods still have significant limitations in modeling motion features: incomplete motion representation: many methods use Euler angles or unit quaternions to describe joint rotation. Euler angles suffer from the well-known "gimbal lock" problem and cannot represent the translational components in motion; while unit quaternions can handle 3D rotation, they can only describe 3 degrees of freedom rotation, failing to uniformly and compactly express the simultaneous rotation and translation (i.e., 6 degrees of freedom helical motion) of human limbs in real-world movement. Insufficient feature discriminative power: existing features (such as joint coordinate sequences) often confuse overall human motion (such as forward and lateral movement) with local relative joint motion (such as knee flexion and extension). Overall motion is greatly influenced by the walking path and environment, and is not an inherent characteristic of the individual; while local relative joint motion more directly reflects the unique characteristics of an individual's physiological structure and movement habits. The failure to separate these two types of information at the kinematic level limits the representational and discriminative power of the features. Sensitive to individual differences: Features are easily influenced by static body shape differences such as individual height and limb length, rather than focusing on the dynamic movement patterns themselves. This affects the generalization ability of the recognition model across different populations. Dual quaternions are a mathematical tool that can simultaneously and efficiently represent rotation and translation in three-dimensional space, while Plück coordinates provide a concise and powerful way to describe lines in three-dimensional space. Currently, combining these two mathematical tools and further decoupling the absolute movement of bones from the relative movement of joints at the kinematic level to construct hierarchical, highly discriminative gait features is the core problem that this invention aims to solve. Summary of the Invention
[0003] This invention aims to provide a hierarchical gait feature extraction and recognition method. By extracting the absolute motion features of bones in the global coordinate system and the relative motion features of joints in the local coordinate system, and effectively fusing them, a more refined and robust representation of individual gait can be achieved, ultimately improving the accuracy of identity recognition. To achieve the above objective, this invention proposes a two-level motion feature modeling framework: First level: Absolute motion modeling of bones. Dual quaternions are used to calculate the 6-DOF rigid body transformation of each independent bone segment between consecutive frames. This feature fully describes the motion trajectory of the bones in the global coordinate system, including the overall movement and posture changes of the human body. Second level: Relative motion modeling of joints. Based on the absolute motion, the relative transformation between adjacent bones is calculated by combining the absolute motion dual quaternions of adjacent bones. This feature filters out the overall human body movement, purely revealing the local motion patterns of joints (such as knees and hips), such as flexion, extension, adduction, and abduction, which are closely related to individual physiological characteristics. Finally, the features from these two levels are aggregated to construct a gait feature sequence with complementary information and strong discriminative power, which is then input into a deep learning model for recognition.
[0004] To achieve the above objectives, the present invention provides the following technical solution: a gait recognition method based on the Plück line, comprising the following steps:
[0005] Step S1: Obtain gait sequence data of the human body, wherein the gait sequence data includes the continuous position coordinates of the main joints of the human body in three-dimensional space;
[0006] Step S2: Based on the gait sequence data, construct a spatial straight-line model of the human skeletal chain using Plück coordinates, where each bone segment is represented as a Plück line vector;
[0007] Step S3: Calculate the absolute rigid body transformation dual quaternion of each bone segment between adjacent frames, which is called the absolute motion feature of the bone; for each joint composed of two adjacent bone segments, calculate its relative transformation dual quaternion based on its corresponding absolute motion feature of the bone, which is called the relative motion feature of the joint.
[0008] Step S4: Aggregate the absolute motion features of all bones and the relative motion features of all joints to construct a gait motion feature sequence;
[0009] Step S5: Input the gait motion feature sequence into a temporal deep learning model for feature encoding and identity recognition.
[0010] Specifically, this can be implemented as follows:
[0011] Step S1: Gait sequence acquisition and preprocessing.
[0012] S1.1: Input video sequence: Obtain a monocular or multi-view video sequence containing human gait.
[0013] S1.2: 3D skeleton reconstruction: Extract the two-dimensional coordinates of human key points in each frame of the image to form a two-dimensional skeleton sequence; based on the two-dimensional skeleton sequence, recover the coordinates of human joints in three-dimensional space in each frame through a three-dimensional pose estimation algorithm or motion structure recovery technology to obtain a three-dimensional skeleton sequence; perform denoising and smoothing filtering preprocessing on the three-dimensional skeleton sequence.
[0014] Step S2: Linear modeling of the skeleton space based on Plück coordinates.
[0015] S2.1: Limb definition: Divide the human skeleton model into M rigid limb segments, each limb being formed by two adjacent three-dimensional joints.
[0016] S2.2: Plück Coordinate Calculation: For each frame K in the 3D skeleton sequence, calculate the Plück coordinates of each limb segment, representing the human limbs connecting joints as Plück lines. For a 3D skeleton sequence... and The defined line segment, whose Plück line is represented as , where the direction vector Moment vector Plück coordinates provide a complete, minimal, and unique representation of a straight line in three-dimensional space. To further eliminate individual differences in body shape, Plück coordinates can be normalized. Through this step, we obtain Plück coordinates that describe the spatial geometry of each limb in each frame.
[0017] Step S3: Dual quaternion representation of limb motion between adjacent frames.
[0018] S3.1: Establish motion model: For two consecutive adjacent frames (frame k and frame k+1), calculate the rigid body motion (rotation and translation) of the same limb line segment from frame k to frame k+1.
[0019] S3.2: Construction of Dual Quaternions: Combining the Plück coordinates obtained in step S2 with dual quaternion theory. A rigid body motion can be represented by a dual quaternion. ,That The rotation is represented by a unit quaternion. It is a quaternion, which is related to both rotation and translation. ,in It is a pure quaternion form of the translation vector.
[0020] The transformation of a straight line represented by Plück coordinates under rigid body motion represented by dual quaternions is linear and can be efficiently accomplished through multiplication of dual quaternions. By solving an optimization problem, the optimal dual quaternion that best describes the motion of the limb from frame k to frame k+1 can be calculated.
[0021] Step S4: Dual quaternion representation of adjacent bone motion.
[0022] The calculation of relative joint motion characteristics specifically involves: for a joint connected by bones a and b, obtaining its absolute motion characteristics within the same time interval. and Calculate the relative motion characteristics of the joint. ,in yes The reverse.
[0023] Step S4: Gait feature sequence construction.
[0024] For all consecutive frame pairs in the video sequence, repeat step S3 to calculate a series of dual quaternions for each limb. Arrange the dual quaternions of all frames and all limbs in temporal and limb order to form a high-dimensional, dynamic gait feature sequence. This sequence compactly encodes the complete motion information of all limbs in three-dimensional space during human walking.
[0025] Step S5: Feature encoding and classification recognition.
[0026] S5.1: Feature Sequence Encoding: Input the long sequence dual quaternion features obtained in step S4 into a temporal feature encoder. The purpose of this encoder is to capture the long-term temporal dependencies of gait motion and output a fixed-length, highly discriminative gait feature vector.
[0027] S5.2: Classifier Training and Recognition: The obtained gait feature vectors are input into a classifier for pedestrian identification. During the training phase, labeled gait data is used to train the entire network (feature extractor + encoder + classifier) end-to-end. In the recognition phase, the gait sequence to be recognized is input, its feature vector is extracted, and identification is completed by comparing its similarity with registered feature vectors in the database.
[0028] Compared with the prior art, the beneficial effects of the present invention are:
[0029] First, the algorithm exhibits viewpoint invariance and body shape adaptability. By using Plück coordinates to essentially describe the straight lines of limbs in three-dimensional space, and combining this with dual quaternions to uniformly represent the helical motion of rigid bodies, the feature extraction process is entirely based on the geometric characteristics of three-dimensional space, fundamentally overcoming the sensitivity of two-dimensional methods to viewpoint changes. Simultaneously, normalization effectively eliminates the influence of differences in limb length among different individuals.
[0030] Secondly, the algorithm achieves optimization of motion representation. Dual quaternions compactly encode rotation and translation information with eight parameters, avoiding the gimbal lock problem of Euler angles and providing a mathematically more stable and complete motion description. This representation method maintains the integrity of motion while ensuring computational efficiency. Attached Figure Description
[0031] Figure 1 This is a flowchart illustrating the gait recognition process based on the Plück line of this invention. Figure 2 This is a schematic diagram of the Plück coordinate geometry of the present invention. Detailed Implementation
[0032] 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.
[0033] Please see Figure 1-2 This invention provides a technical solution for a gait recognition method based on the Plück line: the method includes the following steps:
[0034] Step S1: Obtain gait sequence data of the human body, wherein the gait sequence data includes the continuous position coordinates of the main joints of the human body in three-dimensional space;
[0035] Step S2: Based on the gait sequence data, construct a spatial straight-line model of the human skeletal chain using Plück coordinates, where each bone segment is represented as a Plück line vector;
[0036] In step S2, the construction of the spatial straight line model using Plück coordinates specifically involves: for the connection of the first joint... Second joint The skeleton, its Plück coordinate vector express , where the direction vector Moment vector ,symbol Represents the cross product of vectors;
[0037] Following step S2, the normalization of the Plück coordinates is also performed: Calculation This is to eliminate the impact of individual differences in limb length.
[0038] Step S3: Calculate the absolute rigid body transformation dual quaternion of each bone segment between adjacent frames, which is called the absolute motion feature of the bone; for each joint composed of two adjacent bone segments, calculate its relative transformation dual quaternion based on its corresponding absolute motion feature of the bone, which is called the relative motion feature of the joint.
[0039] In step S3, calculating the absolute motion features of the skeleton specifically involves: for a skeleton segment, its Plück coordinates in the k-th frame are... The Plück coordinates in the (k+1)th frame are Calculate dual quaternions To satisfy ,in This represents dual quaternion multiplication. yes conjugate, This refers to the absolute motion characteristics of the skeleton from frame k to frame k+1.
[0040] In step S3, calculating the relative motion characteristics of the joint specifically involves: for a joint connected by bones a and b, obtaining its absolute motion characteristics within the same time interval. and Calculate the relative motion characteristics of the joint. ,in yes The reverse.
[0041] Step S4: Aggregate the absolute motion features of all bones and the relative motion features of all joints to construct a gait motion feature sequence;
[0042] In step S4, constructing a standardized gait motion feature sequence includes: aggregating the absolute motion features and relative motion features in chronological order to construct an initial feature tensor;
[0043] The dynamic time warping algorithm is used to align all sequences to a standard length in the time dimension; the aligned feature tensors are then Z-score normalized.
[0044] Step S5: Input the gait motion feature sequence into a temporal deep learning model for feature encoding and identity recognition.
[0045] The classifier in step S5 is a temporal deep learning model, specifically a long short-term memory network, a temporal convolutional network, or a Transformer encoder.
[0046] A gait recognition system for implementing a gait recognition method based on the Plück line includes: a data acquisition module for acquiring gait sequence data of a human body; a data processing module for executing steps S2 to S5 to construct a gait feature vector; and a recognition module for storing a gait feature template library and running the classifier to complete identity recognition.
[0047] The steps of the above method can be specifically implemented as follows:
[0048] Step S201: Data preprocessing and 3D skeleton sequence reconstruction (corresponding to S1);
[0049] S201.1: Two-dimensional skeleton extraction: Process each frame of the input video to extract the coordinates and confidence scores of 25 two-dimensional human key points, forming a two-dimensional skeleton sequence.
[0050] S201.2: 3D Skeleton Reconstruction: A learning-based monocular 3D pose estimation model is adopted. This model takes a 2D skeleton sequence as input and uses a spatiotemporal convolutional network to directly regress the 3D skeleton sequence in the camera coordinate system.
[0051] S201.3: Preprocessing: The reconstructed 3D skeleton sequence is subjected to mean filtering to smooth out noise caused by jitter in pose estimation. Subsequently, coordinate system normalization is performed: the pelvic keypoints are used as the origin of the coordinate system, and all skeletons are scaled to a standard scale based on average height.
[0052] Step S202: Geometric modeling of limb space based on Plück coordinates (corresponding to S2);
[0053] S202.1: Limb Segmentation: Based on the standard human body topology, define M key rigid limb segments. For each frame, directly obtain the three-dimensional coordinates of the start and end points of each limb based on the joint coordinates.
[0054] S202.2: Plück coordinate calculation and normalization. For a limb in the k-th frame, its starting point is... The destination is Calculate the direction vector. sum of moments ,in , Normalization: Normalizing the Plück coordinates to eliminate the influence of the absolute length of human limbs yields the following result. This makes the Plück coordinates independent of the actual length of the limb, retaining only its direction and relative position in space.
[0055] Step S203: Calculation of features of dual-level dual quaternions (corresponding to S3);
[0056] This step involves core calculations and is divided into two levels.
[0057] Level 1: Calculation of absolute motion features of the skeleton. Taking skeleton a as an example, calculate the dual quaternions of the rigid body transformation from frame k to frame k+1. A point cloud registration method based on singular value decomposition is adopted. The calculation method is as follows:
[0058] 1. Construct the point cloud. Treat the two endpoints of skeleton a as a rigid body; the source point cloud is... The target point cloud is .
[0059] 2. Calculate the centroid of the limb's point cloud in both frames.
[0060] 3. Decentralization
[0061] 4. Calculate the covariance matrix :
[0062] 5. Singular Value Decomposition (SVD):
[0063] 6. Calculate the rotation matrix :
[0064] 7. Rotate the matrix Convert to unit quaternion And calculate the translation vector
[0065] 8. Constructing the dual quaternion: The dual quaternion can be obtained as follows The same process is used for calculation. .
[0066] Level 2: Calculation of relative joint motion characteristics.
[0067] Calculate the local motion features of the joints connecting bone a and bone b within the same time interval (from frame k to frame k+1). Calculate the relative motion of the joints .
[0068] Step S204: Standardization construction of gait motion feature sequences (corresponding to S4);
[0069] Feature aggregation: For a gait cycle containing N frames, N-1 consecutive frame pairs are obtained. For each frame pair k, the following can be obtained: Absolute motion characteristics of individual bones and Relative motion characteristics of each joint They are concatenated along the feature dimension, with each frame having a feature dimension of 1. Construct initial 3D feature tensor This feature tensor completely records the movement trajectories of all limbs and joints within a gait cycle.
[0070] Temporal standardization: Dynamic Time Warping (DTW) is employed. A sequence of the most representative length from the training set is selected as a template. For each sequence, the optimal alignment path between it and the template is calculated. Based on the alignment path, linear interpolation is used to standardize the length of all sequences to a standard length L, resulting in a uniformly lengthped feature tensor. .
[0071] Feature standardization: Calculate the mean of each feature dimension on the training set. and standard deviation Transform all the data: The final standardized gait motion feature tensor is obtained. It can be directly input into the model for training and recognition.
[0072] Step S205: Feature encoding and identity recognition (corresponding to S5);
[0073] Model input: Standardized feature tensor Remodeling into a two-dimensional sequence ,in This two-dimensional sequence serves as the direct input to the temporal deep learning model.
[0074] Model Training and Feature Encoding: A Bidirectional Long Short-Term Memory (Bi-LSTM) network is preferred as the temporal encoder. During the training phase, the model is trained using a dataset of 3D skeleton gait sequences containing labeled identities. After training, the encoder can map the input gait feature sequence into a highly discriminative fixed-dimensional feature vector.
[0075] Identity Registration: Collect several gait sequences for each user to be registered, extract their respective feature vectors, calculate their average value, and generate the user's gait identity template. And store it in the registration database.
[0076] Online identification and verification: Includes two working modes:
[0077] (Identity Recognition - Closed Set Recognition): For identity recognition tasks, the gait sequence to be recognized is input into a feature encoder to obtain its feature vector. Calculate from the pre-registered gait feature library. The system calculates the cosine similarity between the template vectors and all template vectors in the library, and finally outputs the identity corresponding to the template with the highest similarity as the recognition result.
[0078] (Authentication - Open Set Confirmation): For authentication tasks, given a claimed identity and its corresponding gait sequence, the system needs to verify whether the claim is true. First, the system retrieves the registered feature vector corresponding to the claimed identity from the feature library. Then calculate the feature vector of the input sequence. and Similarity score Finally, the score is compared with a preset decision threshold. Comparison: If If the claimed identity is accepted, the verification will pass; otherwise, the verification will fail.
[0079] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A gait recognition method based on the Plück line, characterized in that: The method includes the following steps: Step S1: Obtain gait sequence data of the human body, wherein the gait sequence data includes the continuous position coordinates of the main joints of the human body in three-dimensional space; Step S2: Based on the gait sequence data, construct a spatial straight-line model of the human skeletal chain using Plück coordinates, where each bone segment is represented as a Plück line vector; Step S3: Calculate the absolute rigid body transformation dual quaternion of each bone segment between adjacent frames, which is called the absolute motion feature of the bone; for each joint composed of two adjacent bone segments, calculate its relative transformation dual quaternion based on its corresponding absolute motion feature of the bone, which is called the relative motion feature of the joint. Step S4: Aggregate the absolute motion features of all bones and the relative motion features of all joints to construct a gait motion feature sequence; Step S5: Input the gait motion feature sequence into a temporal deep learning model for feature encoding and identity recognition.
2. The gait recognition method based on the Plück line according to claim 1, characterized in that: In step S2, the construction of the spatial straight line model using Plück coordinates specifically involves: for connecting the first joint... Second joint The skeleton, its Plück coordinate vector express , where the direction vector Moment vector ,symbol This represents the cross product of vectors.
3. The gait recognition method based on the Plück line according to claim 1, characterized in that: Following step S2, the normalization of the Plück coordinates is also performed: Calculation This is to eliminate the impact of individual differences in limb length.
4. The gait recognition method based on the Plück line according to claim 1, characterized in that: In step S3, calculating the absolute motion features of the skeleton specifically involves: for a skeleton segment, its Plück coordinates in the k-th frame are... The Plück coordinates in the (k+1)th frame are Calculate dual quaternions To satisfy ,in This represents dual quaternion multiplication. yes conjugate, This refers to the absolute motion characteristics of the skeleton from frame k to frame k+1.
5. The gait recognition method based on the Plück line according to claim 1, characterized in that: In step S3, calculating the relative motion characteristics of the joint specifically involves: for a joint connected by bones a and b, obtaining its absolute motion characteristics within the same time interval. and Calculate the relative motion characteristics of the joint. ,in yes The reverse.
6. The gait recognition method based on the Plück line according to claim 1, characterized in that: In step S4, constructing a standardized gait motion feature sequence includes: aggregating the absolute motion features and relative motion features in chronological order to construct an initial feature tensor; The dynamic time warping algorithm is used to align all sequences to a standard length in the time dimension; the aligned feature tensors are then Z-score normalized.
7. The gait recognition method based on the Plück line according to claim 1, characterized in that: The classifier in step S5 is a temporal deep learning model, specifically a long short-term memory network, a temporal convolutional network, or a Transformer encoder.
8. A gait recognition system for implementing the method as described in any one of claims 1-7, characterized in that, include: The data acquisition module is used to acquire human gait sequence data; The data processing module is used to execute steps S2 to S5 to construct gait feature vectors; The recognition module is used to store the gait feature template library and run the classifier to complete identity recognition.