Gait recognition method and system based on local and global difference modeling network

Abstract

The application discloses a gait recognition method and system based on local and global difference modeling networks, relates to the field of gait recognition, and comprises the following steps: acquiring a gait contour sequence and environment parameters of a target object; processing the gait contour sequence according to the environment parameters to generate local limb motion trajectory segments and a global gait cycle sequence; performing feature extraction on the local limb motion trajectory segments and the global gait cycle sequence to obtain a local difference feature vector and a global difference feature vector; generating a fusion difference feature according to the local difference feature vector and the global difference feature vector; performing dynamic weight distribution on the fusion difference feature according to the environment parameters to obtain an identity discrimination feature; selecting a matching strategy according to the environment parameters, performing similarity matching on the identity discrimination feature and a pre-stored gait feature library according to the matching strategy, and outputting an identity determination result of the target object; and the application can overcome complex environmental interference and improve the stability of feature extraction.

Description

Technical Field

[0001] This invention relates to the field of gait recognition, specifically to a gait recognition method and system based on local and global difference modeling networks. Background Technology

[0002] Gait recognition, as an emerging biometric identification technology, has broad application prospects in fields such as security monitoring and identity authentication due to its advantages such as non-contact and long-distance recognition. However, existing gait recognition methods still face many challenges and have some obvious shortcomings:

[0003] In real-world applications, complex environmental factors such as changes in lighting and perspective significantly impact the extraction and recognition of gait features. Traditional methods perform poorly under varying environmental parameters, particularly in low-light conditions where the quality of gait images acquired by image acquisition devices deteriorates, leading to inaccurate feature extraction. Furthermore, the appearance of gait changes considerably when pedestrians walk from different perspectives, drastically reducing the accuracy of models trained on fixed perspectives. These issues severely limit the effectiveness of gait recognition technology in real-world scenarios.

[0004] Existing methods have limitations in feature extraction. Many methods focus on extracting global gait features while neglecting the importance of local details. Although global features can reflect the overall pattern of gait, they are insufficient in capturing some subtle, individual-specific local features. These local features play a crucial role in distinguishing similar gaits, and the lack of effective extraction of them limits the accuracy of gait recognition. Furthermore, existing methods lack effective feature fusion mechanisms, making it difficult to achieve synergistic optimization of local and global features.

[0005] To address the aforementioned problems, existing technologies urgently need improvement. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a gait recognition method and system based on local and global difference modeling networks.

[0007] To achieve the above objectives, the technical solution of the present invention is as follows:

[0008] In a first aspect, the present invention discloses a gait recognition method based on local and global difference modeling networks, comprising the following steps:

[0009] Obtain the gait contour sequence and environmental parameters of the target object;

[0010] The segmentation granularity is adjusted according to the environmental parameters, and the gait contour sequence is subjected to spatiotemporal segmentation to generate local limb movement trajectory fragments and global gait cycle sequences.

[0011] The local limb movement trajectory segment is subjected to frequency domain feature transformation to obtain a local difference feature vector, and the global gait cycle sequence is subjected to spatial alignment encoding to obtain a global difference feature vector;

[0012] The local difference feature vector and the global difference feature vector are subjected to cross-scale interactive calculation to generate a fused difference feature with spatiotemporal consistency.

[0013] Based on the environmental parameters, the fusion difference features are dynamically weighted to obtain optimized identity discrimination features;

[0014] A matching strategy is selected based on the environmental parameters; the matching strategy includes three-level full feature matching and lower limb feature-priority matching.

[0015] According to the corresponding matching strategy, the identity discrimination features are matched with the pre-stored gait feature library for similarity, and the identity determination result of the target object is output.

[0016] Secondly, this invention discloses a gait recognition system based on local and global difference modeling networks, comprising:

[0017] The data acquisition module is used to acquire the gait contour sequence and environmental parameters of the target object;

[0018] The data processing module is used to adjust the segmentation granularity according to the environmental parameters, perform spatiotemporal segmentation processing on the gait contour sequence, and generate local limb movement trajectory fragments and global gait cycle sequences.

[0019] The feature extraction module is used to perform frequency domain feature transformation on the local limb movement trajectory segment to obtain a local difference feature vector, and to perform spatial alignment encoding on the global gait cycle sequence to obtain a global difference feature vector;

[0020] The feature fusion module is used to perform cross-scale interactive calculations on the local difference feature vector and the global difference feature vector to generate fused difference features with spatiotemporal consistency; and to dynamically assign weights to the fused difference features according to the environmental parameters to obtain optimized identity discrimination features.

[0021] The strategy selection module is used to select a matching strategy based on the environmental parameters; the matching strategy includes three-level full feature matching and lower limb feature-priority matching.

[0022] The feature matching module is used to perform similarity matching between the identity discrimination features and the pre-stored gait feature library according to the corresponding matching strategy, and output the identity determination result of the target object.

[0023] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0024] 1. By monitoring light intensity and viewing angle in real time, the spatiotemporal segmentation granularity is dynamically adjusted (e.g., increasing upper limb sampling density under strong light and focusing on lower limb features under weak light / side view), effectively overcoming interference from complex environments and improving the stability of feature extraction;

[0025] 2. By fusing data from visible light cameras (RGB images) and depth sensors (3D skeleton points), depth data is prioritized to compensate for contour information under low light conditions. Kinematic constraints are used to verify and eliminate outliers, ensuring the integrity and rationality of the input data.

[0026] 3. Short-time Fourier transform is used to extract local joint frequency domain features, combined with global bone point density gradient analysis, and local and global features are fused through environmental awareness attention mechanism to enhance spatiotemporal consistency expression ability;

[0027] 4. Automatically switch matching modes based on environmental parameters: In strong light / frontal view, enable three-level full feature matching; in weak light / side view, enable lower limb priority matching to suppress features such as upper limbs and trunk that are easily affected by view / occlusion. When the view is off or the lighting is insufficient, the anti-interference characteristics of lower limb movement are used to improve recognition robustness. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0029] Figure 1 This is an overall method block diagram of Embodiment 1 of the present invention;

[0030] Figure 2 This is a flowchart of the method according to Embodiment 1 of the present invention;

[0031] Figure 3 This is an overall system block diagram of Embodiment 2 of the present invention. Detailed Implementation

[0032] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. 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] Application Overview: In traditional gait recognition methods, changes in environmental parameters lead to biases in the feature extraction process, insufficient spatiotemporal correlation between global and local features, and a lack of adaptability in multi-scale feature fusion under dynamic scenes. Contour extraction algorithms under complex lighting conditions cannot eliminate pixel distortion in low-light areas, viewpoint shifts cause abnormal spatial distribution dispersion of skeletal points, spatiotemporal segmentation processing lacks a correlation mechanism with physical environment parameters, and the cross-scale feature interaction process does not introduce an environmental perception mechanism, resulting in representation errors in the dynamic weight allocation of fused differential features.

[0034] For example, gait recognition devices deployed in security monitoring systems synchronously collect data via visible light cameras and depth sensors. When ambient light intensity decreases and the acquisition viewpoint deviates from the normal direction by a certain angle, the gait sequence output by the contour extraction module shows broken lower limb joint trajectories and misaligned upper limb motion phase markers. The spatiotemporal segmentation processing module uses fixed granularity parameters, leading to oversampling of the upper limb chain motion trajectory in strong light environments, resulting in noise interference, and undersampling of the lower limb chain motion trajectory in low light environments, resulting in the loss of key gait phases. The local difference feature vector output by the frequency domain feature transformation module and the global difference feature vector generated by the spatial alignment encoding module have a scale mismatch. During cross-scale interaction, no mapping relationship between environmental parameters and feature weights is established, ultimately causing cosine similarity calculation errors in identity discrimination features during dynamic matching.

[0035] If the above issues are not addressed, the identity determination module will be unable to accurately distinguish individual differences within similar gait patterns, leading to an increased false recognition rate across different environments. Static features stored in the gait feature library cannot adapt to feature drift in dynamic environments, and asynchronous timestamps from multi-source sensor data cause phase misalignment in feature fusion, significantly reducing the system's robustness in complex scenarios. Feature representation mismatch caused by changes in perspective during identity authentication will lead to security vulnerabilities, limiting the real-time detection capability of abnormal gait behavior in monitored scenarios.

[0036] To address the aforementioned challenges, this application first considers establishing a dynamic correlation mechanism between environmental parameters and the data processing process. To resolve the issue of feature inaccuracies caused by changes in illumination and viewing angle, it attempts to synchronize sensor data with environmental parameters via timestamps, dynamically adjusting the feature extraction strategy based on these parameters. To address the insufficient spatiotemporal correlation between global and local features, it explores introducing a cross-scale interaction mechanism after frequency domain transformation and spatial encoding, enabling local motion trajectories and global periodic sequences to form complementary features. For the dynamic weight allocation bias problem, a gated fusion method based on the environmental parameter encoding vector is designed, controlling the feature fusion ratio through a trainable weight matrix. Finally, it chooses to link the spatiotemporal segmentation granularity with environmental parameters, establishing an adaptive adjustment mechanism for the segmentation mode, combining cross-scale feature interaction and dynamic weight allocation to form a complete technical chain.

[0037] Example 1:

[0038] like Figure 1-2 As shown, the gait recognition method based on local and global difference modeling networks includes the following steps: acquiring the gait contour sequence and environmental parameters of the target object; adjusting the segmentation granularity according to the environmental parameters, performing spatiotemporal segmentation on the gait contour sequence to generate local limb movement trajectory segments and a global gait cycle sequence; performing frequency domain feature transformation on the local limb movement trajectory segments to obtain local difference feature vectors, and performing spatial alignment encoding on the global gait cycle sequence to obtain global difference feature vectors; performing cross-scale interactive calculation on the local difference feature vectors and global difference feature vectors to generate fused difference features with spatiotemporal consistency; dynamically assigning weights to the fused difference features according to the environmental parameters to obtain optimized identity discrimination features; selecting a matching strategy according to the environmental parameters; the matching strategy includes three-level full feature matching and lower limb feature-priority matching; performing similarity matching between the identity discrimination features and a pre-stored gait feature library according to the corresponding matching strategy, and outputting the identity determination result of the target object.

[0039] This application further proposes to acquire gait contour sequences through a multi-source sensing device, which includes at least a visible light camera and a depth sensor; to acquire RGB gait image sequences through the visible light camera, while simultaneously acquiring the motion trajectory of three-dimensional skeletal points through the depth sensor; when extracting contours from the RGB gait image sequences, an adaptive illumination compensation algorithm is used to eliminate pixel distortion in low-light areas; and to perform kinematic constraint verification on the motion trajectory of three-dimensional skeletal points, eliminating abnormal data points that exceed the human joint activity threshold.

[0040] The multi-source sensing device employs a synchronous triggering mechanism between the visible light camera and the depth sensor to ensure that the timestamps of the RGB image and the skeletal point trajectory are aligned. An adaptive illumination compensation algorithm dynamically adjusts the gamma correction coefficients based on the image histogram distribution to locally enhance brightness in low-light areas. Kinematic constraint verification detects abrupt changes in the skeletal point trajectory that do not conform to the laws of human movement by establishing a joint angle change rate model.

[0041] Specifically, the visible light camera captures RGB image sequences at a rate of 30 frames per second. In low-light environments, an adaptive illumination compensation algorithm divides the image into 8×8 pixel blocks, calculates the average brightness of each block, and applies a correction curve with a preset gamma value to areas below 50 lux to enhance details in dark areas. The depth sensor outputs the 3D coordinates of skeletal points at the same frame rate. When a knee flexion angle exceeding 150 degrees or an ankle rotation speed exceeding 120 degrees per second is detected, it is identified as an abnormal data point and removed. Through a multi-source data complementarity mechanism, depth sensor data is preferentially used to supplement the missing contour information in the visible light image under low-light conditions. At the same time, physical constraints ensure the physiological rationality of the motion trajectory, providing high-quality input for subsequent spatiotemporal segmentation.

[0042] Through the above technical solutions, this application achieves effective processing of gait images under complex lighting conditions, improving the accuracy of contour extraction. Simultaneously, kinematic constraint verification ensures the validity of skeletal point data, laying the foundation for subsequent feature extraction. This enhances the robustness and adaptability of the gait recognition system under different environmental conditions.

[0043] In some of the solutions described above in this application, the methods for obtaining environmental parameters suffer from insufficient timeliness and inaccurate parameter classification, resulting in the inability to accurately correlate the influencing factors of environmental changes during gait feature extraction. For example, the lack of graded classification of illumination intensity may prevent subsequent algorithms from effectively distinguishing feature processing modes under different illumination conditions; the failure to classify viewpoints according to azimuth intervals may lead to a decrease in the accuracy of spatial alignment encoding; and the asynchronous timestamps of parameters and gait data may cause temporal misalignment during data processing.

[0044] This application further proposes a process for acquiring environmental parameters, including light intensity and acquisition viewing angle; real-time measurement of light intensity using a photosensitive sensor built into a visible light camera; calculation of the angle between the centroid of the target object and the normal to the acquisition plane using a depth sensor, which serves as the acquisition viewing angle; division of light intensity into three levels—strong light, normal light, and weak light—based on a preset light intensity threshold; division of the acquisition viewing angle into two azimuth ranges—frontal view and side view—based on a preset angle threshold; and establishment of a timestamp synchronization mechanism between environmental parameters and gait contour sequences to ensure the timeliness consistency of parameter association.

[0045] Real-time measurement of light intensity is achieved through a photosensitive sensor, whose response frequency matches the camera frame rate to ensure the real-time nature of the light data for each frame. The acquisition viewpoint angle is obtained through geometric calculation of the angle between the centroid and the normal to the acquisition plane. Specifically, after constructing a plane equation from the 3D coordinate data output by the depth sensor, the angle between the centroid coordinate projection and the normal is calculated. Light intensity thresholds are defined using fixed numerical ranges, for example, greater than 1000 lux for strong illumination, 300-1000 lux for normal illumination, and less than 300 lux for weak illumination. Angle thresholds use a preset 30 degrees as the dividing point between the frontal and lateral viewpoints; an angle less than 30 degrees is considered a frontal viewpoint, otherwise a lateral viewpoint. The timestamp synchronization mechanism is triggered by a hardware clock to align the environmental parameter acquisition time with the gait contour sequence generation time, with errors controlled within milliseconds.

[0046] Specifically, the photosensor is integrated into the camera module, directly reading the voltage signal output by the sensor and converting it into a light intensity value, avoiding interference from the external environment. The depth sensor calculates the centroid position using the 3D coordinates of the skeletal points, and uses the normal vector of the acquisition plane as a reference to calculate the cosine value of the included angle using the vector dot product formula. Light intensity level classification is achieved by comparing real-time measured values ​​with preset threshold ranges; for example, a measured value of 500 lux is determined to be a normal light level. Viewpoint azimuth range classification is achieved by comparing the calculated included angle with a 30-degree threshold; for example, an included angle of 45 degrees is determined to be a side view. Timestamp synchronization is achieved through a synchronization signal trigger mechanism. When the camera and depth sensor acquire data, the central processing unit records a unified system timestamp, ensuring strict temporal correspondence between environmental parameters and gait data in subsequent processing. Thus, the quantitative classification of light intensity and viewpoint provides accurate parameter input for subsequent dynamic adjustment of segmentation granularity and matching strategies, and the timestamp synchronization mechanism eliminates delay errors in data association, ensuring the spatiotemporal consistency of feature extraction.

[0047] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0048] Environmental parameters include light intensity and acquisition angle. Light intensity is measured in real time using a photosensitizer built into the visible light camera. The acquisition angle is obtained by calculating the angle between the target object's centroid and the normal to the acquisition plane using a depth sensor. Light intensity is divided into three levels—strong light, normal light, and weak light—based on preset light intensity thresholds. The acquisition angle is divided into two azimuth ranges—frontal view and side view—based on preset angle thresholds. A timestamp synchronization mechanism is established between environmental parameters and gait contour sequences to ensure the timeliness consistency of parameter association.

[0049] Specifically, the built-in photosensor of the visible light camera collects light intensity data every 100 milliseconds. The depth sensor calculates the angle between the centroid of the target object and the normal to the acquisition plane every 50 milliseconds. Light intensity is divided into three levels: 0-100 lux for weak light, 100-1000 lux for normal light, and above 1000 lux for strong light. The acquisition viewing angle is divided into two azimuth ranges: 0-30 degrees for a frontal view and 30-90 degrees for a side view. The timestamp synchronization mechanism between environmental parameters and gait contour sequences uses a timestamp alignment method, matching the acquisition timestamp of the environmental parameters with the timestamp of the most recent gait contour sequence frame.

[0050] Through the above technical solution, this application achieves accurate acquisition and quantification of environmental parameters. This allows for dynamic adjustment of gait feature extraction and matching strategies based on real-time lighting and viewing conditions, improving the adaptability of the gait recognition system to complex environments. Furthermore, by establishing a time synchronization mechanism between environmental parameters and gait data, the consistency between environmental factors and gait data during feature extraction is ensured, avoiding feature distortion caused by lag in environmental parameters.

[0051] In some of the solutions described above in this application, environmental parameters are divided into different level ranges, but it is not clear how to adjust the granularity of spatiotemporal segmentation based on different environmental parameters. This makes it difficult to balance the contradiction between the integrity of local feature extraction and data redundancy under complex environmental conditions. For example, when gait contour sequences collected under low light conditions are subject to noise interference, using a fixed segmentation mode may result in excessively high sampling density of local limb movement trajectories, introducing invalid feature components.

[0052] This application further proposes dividing the gait contour sequence into N phase intervals, each phase interval corresponding to the spatiotemporal segmentation parameters of the foot contact, leg swing, and takeoff phases; within a single phase interval, the human body is divided into three local regions—upper limb chain, trunk chain, and lower limb chain—using a kinematic chain model; when sampling the motion trajectory for each local region, a time window sliding mechanism is used to ensure that the overlap rate of adjacent trajectory segments is not lower than a preset threshold; the process of adjusting the segmentation granularity according to environmental parameters includes: using a fine-grained segmentation mode under strong light to increase the sampling density of the upper limb chain motion trajectory; using a coarse-grained segmentation mode under weak light to retain the sampling data of the lower limb chain motion trajectory; and retaining the sampling data of the lower limb chain motion trajectory from a side view.

[0053] The phase intervals are divided based on the biomechanical characteristics of the human gait cycle, with each phase interval corresponding to a specific kinematic parameter threshold. The kinematic chain model abstracts the human body as a multi-rigid-body system, establishing the kinematic relationships between each chain segment through joint constraint equations. The time window sliding mechanism uses a variable window length, with the window movement step size dynamically adjusted according to gait frequency, and a preset overlap rate threshold set to P1%. The segmentation granularity adjustment module has a built-in strategy mapping table for different combinations of environmental parameters. For example, when the light intensity level is low and the viewing angle is a side view, the lower limb chain trajectory preservation mode is automatically triggered.

[0054] Specifically, under strong light conditions, the RGB images captured by the visible light camera are of high quality. The fine-grained segmentation mode increases the sampling frequency of the shoulder joint swing trajectory by shortening the sampling interval of the upper limb chain, for example, adjusting the sampling interval of the upper limb chain from 10ms per frame to 5ms. Under low light conditions, the skeletal point data captured by the depth sensor is less affected by noise. The coarse-grained segmentation mode adjusts the sampling interval of the lower limb chain to 20ms, while discarding trajectory points in the upper limb chain with confidence levels below a preset confidence threshold. When the viewpoint is a side view, the projection of the upper limb chain's motion trajectory in the 3D coordinate system exhibits significant deformation. In this case, the lower limb chain data is retained, and the knee joint angle change curve is reconstructed using an inverse kinematics algorithm. The time window sliding mechanism generates overlapping regions between adjacent trajectory segments, and the trajectory data in the overlapping regions is smoothed using a Kalman filter to eliminate trajectory abrupt changes caused by sensor jitter.

[0055] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0056] The gait profile sequence was divided into eight phase intervals, each corresponding to the spatiotemporal segmentation parameters of the foot strike, leg swing, and takeoff phases. Within a single phase interval, the human body was divided into three local regions—upper limb chain, trunk chain, and lower limb chain—using a kinematic chain model. When sampling the motion trajectory for each local region, a time window sliding mechanism was employed to ensure that the overlap rate of adjacent trajectory segments was no less than 80%.

[0057] The process of adjusting the segmentation granularity based on environmental parameters includes: using a fine-grained segmentation mode under strong lighting conditions with an intensity greater than 1000 lux, increasing the sampling density of upper limb chain motion trajectories to 60 frames per second; using a coarse-grained segmentation mode under weak lighting conditions with an intensity less than 100 lux, retaining only the sampling data of lower limb chain motion trajectories, and reducing the sampling density to 30 frames per second; and similarly, retaining only the sampling data of lower limb chain motion trajectories when the angle between the acquisition viewpoint and the frontal direction is greater than 45 degrees.

[0058] Through the above technical solution, this application can adaptively adjust the granularity of gait feature extraction according to different environmental conditions. Under strong lighting conditions, it can fully utilize detailed information to improve recognition accuracy; under unfavorable conditions such as weak lighting or side view, it can reduce the sampling granularity and focus on lower limb features to ensure the stability of feature extraction. This dynamic adjustment mechanism significantly improves the adaptability and robustness of the gait recognition system in complex and changing environments.

[0059] In some of the solutions described above in this application, when extracting features based on local limb motion trajectory segments and global gait cycle sequences after spatiotemporal segmentation, it is difficult to effectively capture the dynamic changes in local joint motion frequencies and accurately characterize the key gradient changes in the spatial distribution of skeletal points, resulting in insufficient feature discrimination.

[0060] This application further proposes to perform short-time Fourier transform on local limb movement trajectory segments to extract the motion frequency distribution features of each joint; to perform dimensionality reduction processing on the motion frequency distribution features using principal component analysis, retaining feature components whose variance contribution rate exceeds a preset threshold; and to divide the global gait cycle sequence into a three-dimensional spatial grid and calculate the density gradient change features of skeletal points within each grid cell. The calculation method for the density gradient change features includes: decomposing the three-dimensional spatial grid into horizontal and vertical slice cells along the motion direction; calculating the spatial distribution dispersion of skeletal points within each slice cell to generate a density gradient change curve; and performing difference calculation on the density gradient change curves of adjacent slice cells to extract curvature abrupt change points as feature markers.

[0061] In this process, the short-time Fourier transform uses a fixed time window to perform spectral analysis on the time-domain signal of local trajectory segments. The window length is set to one-quarter of the gait cycle length, and the overlap rate is no less than P1%. During principal component analysis (PCA) dimensionality reduction, a preset variance contribution rate threshold of P2% is used, retaining the first three principal components. For 3D spatial mesh generation, the space is divided into 20cm×20cm×20cm cubic units along the direction of human movement. Horizontal slice units are cut along the sagittal plane, and vertical slice units are cut along the coronal plane. The spatial distribution dispersion is obtained by calculating the standard deviation of the skeletal point coordinates within the slice unit, and the density gradient change curves are arranged in a time series. The difference calculation uses a first-order forward difference operator, and the curvature abrupt change point is defined as the position where the difference value exceeds twice the standard deviation of the mean.

[0062] Specifically, after a short-time Fourier transform, the frequency distribution characteristics of joint motion segments are decomposed into amplitude and phase spectra. The amplitude spectrum characterizes motion intensity, while the phase spectrum characterizes temporal relationships. Principal component analysis (PCA) uses eigenvalue decomposition of the covariance matrix to select the low-frequency components with the strongest representational ability. Three-dimensional spatial meshing maps the global gait cycle sequence to three-dimensional space; lateral slices capture lateral motion features, and longitudinal slices capture forward and backward motion features. The spatial dispersion of skeletal points reflects the concentration of the motion trajectory, and the density gradient change curve forms a dynamic pattern through the accumulation of time. Differential calculation effectively amplifies the distribution differences between adjacent slices, with curvature abrupt changes corresponding to key turning points in the gait cycle. Through these steps, local joint motion frequency characteristics and global skeletal density gradient characteristics complement each other, enhancing the representational ability of the features.

[0063] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0064] The frequency domain feature transformation process includes performing a short-time Fourier transform on the local limb motion trajectory segment to extract the motion frequency distribution features of each joint. Furthermore, principal component analysis is used to reduce the dimensionality of the motion frequency distribution features, retaining feature components whose variance contribution rate exceeds a preset threshold.

[0065] The global gait cycle sequence is divided into a three-dimensional spatial mesh, and the density gradient variation characteristics of skeletal points within each mesh cell are calculated. The method for calculating the density gradient variation characteristics involves decomposing the three-dimensional spatial mesh into horizontal and vertical slice cells along the motion direction. Specifically, the spatial distribution dispersion of skeletal points within each slice cell is calculated to generate a density gradient variation curve. Then, the density gradient variation curves of adjacent slice cells are differentially calculated, and curvature abrupt change points are extracted as feature markers.

[0066] For example, in practical applications, the gait cycle sequence can be divided into a 10×10×10 three-dimensional grid. For each grid cell, the spatial distribution dispersion of the skeletal points is calculated, forming a 100×100 density gradient matrix. Then, the matrix is ​​decomposed into 10 10×10 slice cells along the direction of motion. The standard deviation of the skeletal point distribution is calculated for each slice cell, resulting in a density gradient change curve composed of 10 standard deviation values. Finally, the density gradient change curves of adjacent slice cells are differentially analyzed, and points with a difference value greater than a preset threshold are extracted as curvature abrupt change points; these abrupt change points are the feature markers.

[0067] Through the above technical solution, this application can effectively extract the frequency domain features of local limb movements and the spatial features of the global gait cycle. Frequency domain feature transformation can capture the periodic patterns of joint movements, while density gradient changes reflect the spatial distribution patterns of the overall gait. This multi-scale feature extraction method improves the expressive power of gait features, enabling the model to simultaneously focus on local details and global patterns, thereby enhancing the accuracy and robustness of gait recognition. Especially in complex environments, such as changes in lighting or viewing angle, this method can still maintain good feature extraction performance, improving the adaptability of the gait recognition system to different scenarios.

[0068] In some of the schemes described above in this application, the impact of environmental changes on the fusion of local and global features during cross-scale interactive computing is not fully modeled, resulting in a lack of adaptability in the allocation of feature weights under different environmental conditions, which affects the discriminative ability of the fused features.

[0069] This application further proposes establishing a dynamic mapping relationship between local and global difference feature vectors, and calculating the contribution weight of each local region to the global features through a spatial attention mechanism. Cross-scale interactive computation also includes an environment-aware attention mechanism: gating coefficients are generated based on the environment parameter encoding vector E, and the calculation process satisfies Aenv=σ(WlL+WgG+We*E), where L is the local difference feature vector, G is the global difference feature vector, Wl, Wg, and We are trainable weight matrices, and σ is the sigmoid activation function. Aenv is used as a dynamic weight to control the fusion ratio of local and global features. The generation method of the environment parameter encoding vector E includes: performing a logarithmic transformation on the illumination intensity to obtain a normalized illumination encoding value; performing sine and cosine encoding on the acquisition viewpoint angle to generate an azimuth feature vector; and concatenating the two types of encoding vectors and then reducing the dimensionality through a fully connected layer to form the environment parameter encoding vector E.

[0070] The dynamic mapping relationship is achieved through a spatial attention mechanism, which calculates the association weight of the motion trajectory segment of each local limb region with the global gait cycle sequence. The environmental awareness attention mechanism applies the gating coefficient Aenv to the feature fusion process, linearly combining local features L, global features G, and environmental parameters E through a trainable weight matrix. The logarithmic transformation of illumination intensity uses a base-10 logarithmic function to compress the light intensity value to the 0-1 interval; the sine and cosine encoding of the viewing angle converts the angle θ into a two-dimensional vector [sinθ, cosθ]. The dimensionality reduction operation of the fully connected layer maps the concatenated encoded vector to a 128-dimensional vector space.

[0071] Specifically, under strong lighting conditions, the normalized lighting encoding value approaches 1. In this case, the gating coefficient Aenv tends to preserve upper limb movement details in the local differential features. Under side-view conditions, the orientation feature vector is encoded into a two-dimensional vector using sine and cosine encoding. This vector is then concatenated with the lighting encoding and passed through a fully connected layer to form the environment parameter encoding vector E. This vector is used to adjust the gating coefficient Aenv through the weight matrix We, making the fusion process focus more on lower limb movement trajectories. The dynamic weights Aenv output by the sigmoid activation function are then multiplied with the feature vector using the Hadamard product to achieve adaptive feature fusion ratio adjustment.

[0072] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0073] A dynamic mapping relationship between local and global difference feature vectors is established, and the contribution weight of each local region to the global features is calculated through a spatial attention mechanism. Cross-scale interactive computation also includes an environment-aware attention mechanism. Gating coefficients are generated based on the environment parameter encoding vector E, and the calculation process satisfies: Aenv = σ(WlL + WgG + We*E), where L is the local difference feature vector, G is the global difference feature vector, Wl, Wg, and We are trainable weight matrices, and σ is the sigmoid activation function. Aenv is used as a dynamic weight to control the fusion ratio of local and global features. The generation method of the environment parameter encoding vector E includes: performing a logarithmic transformation on the illumination intensity to obtain a normalized illumination encoding value; performing sine and cosine encoding on the acquisition viewpoint angle to generate an azimuth feature vector; concatenating the two types of encoding vectors and then reducing the dimensionality through a fully connected layer to form the environment parameter encoding vector E.

[0074] Specifically, a spatial attention module is first constructed, which contains multiple convolutional layers and non-linear activation functions. The local difference feature vector L is input into this module to obtain the attention weights for each local region. Then, these weights are weighted and summed with the global difference feature vector G to obtain the preliminary fused features.

[0075] Furthermore, an environment-aware attention module is constructed. This module first performs a logarithmic transformation on the illumination intensity, mapping the illumination intensity range of 0-10000 lux to the range of 0-1. It then performs sine and cosine encoding on the acquisition viewpoint angle, encoding the angle from 0-360 degrees into a two-dimensional vector. After concatenating the illumination encoding and viewpoint encoding, the vector is reduced in dimensionality through a fully connected layer containing 64 neurons, resulting in a 32-dimensional environment parameter encoding vector E.

[0076] Therefore, L, G, and E are input into the formula Aenv = σ(WlL + WgG + We*E), where Wl, Wg, and We are trainable weight matrices initialized with random values. σ is the sigmoid activation function, mapping the output to the range of 0-1. The resulting Aenv is the dynamic fusion weight for environment perception.

[0077] Finally, the preliminary fusion features are multiplied by Aenv to obtain the final fusion difference features. These features simultaneously incorporate local details, global structure, and the influence of environmental factors.

[0078] Through the above technical solutions, this application achieves adaptive fusion of local and global features and introduces dynamic adjustment of environmental factors. This fusion method can automatically adjust the importance of features according to different environmental conditions, improving the robustness of gait recognition in complex environments. Simultaneously, the spatial attention mechanism highlights the contribution of key local regions, enhancing the model's ability to capture subtle differences. The environmental awareness attention mechanism enables the model to flexibly adjust the feature fusion strategy according to changes in lighting and viewing angle, further improving the accuracy and generalization ability of recognition.

[0079] In some of the solutions described above in this application, when the light intensity is low or the acquisition viewpoint deviates from the frontal viewpoint, the weight allocation of different data sources in the fusion of differential features fails to dynamically adjust according to environmental changes, resulting in insufficient robustness of identity discrimination features in complex environments and affecting the final matching accuracy. For example, under low light conditions, the quality of gait contour sequences acquired by visible light cameras decreases, while the motion trajectories of skeletal points generated by depth sensors have higher reliability, but existing methods have failed to effectively increase their weight ratio; in side-view acquisition scenarios, the visibility of upper limb chain motion features may decrease, but existing methods have not adaptively adjusted feature priorities.

[0080] This application further proposes a dynamic weight allocation process including: when the light intensity is lower than a preset threshold, increasing the weight ratio of the skeletal point motion trajectory generated by the depth sensor in the local differential feature vector; and when the acquisition viewpoint deviates from the positive viewpoint by more than a preset range, enhancing the priority of the upper limb chain motion features in the global differential feature vector.

[0081] The preset threshold is determined based on the inflection point of the matching error curve between skeletal point trajectory data and contour image features under different illumination intensities. Weight ratio adjustment is achieved by modifying the scaling factor of the feature vector, specifically multiplying the skeletal point trajectory feature components by a preset scaling factor. The viewing angle deviation range is defined as triggering priority adjustment when the angle between the acquisition plane and the normal exceeds θ1 degrees, using a feature vector sorting algorithm to improve the ranking position of the upper limb chain features in the global features by 2-3 places. These adjustments are triggered by real-time monitoring of the environmental parameter encoding vector E, where the illumination intensity parameter is updated every 200 milliseconds and the viewing angle parameter is updated every 500 milliseconds.

[0082] Specifically, under low-light conditions, the ambient light intensity is continuously monitored by a photosensor. When the light intensity value is detected to be below a preset threshold for three consecutive sampling periods, the data processing module automatically activates the depth sensor data weight enhancement mode. At this time, the skeletal point motion trajectory component in the local difference feature vector is multiplied by a preset weight coefficient, while the visible light contour feature component retains its original coefficient. In scenes with viewpoint deviation, when the depth sensor detects that the angle between the centroid of the target object and the normal of the acquisition plane reaches θ2 degrees, the upper limb chain motion features in the global difference feature vector are reordered to the top three of the feature sequence, and their participation in the feature matching stage is enhanced through a spatial attention mechanism. For example, in the gait cycle acquired from a side viewpoint, the upper limb swing amplitude is difficult to capture accurately due to perspective distortion. In this case, the stable features of the lower limb chain motion trajectory are used first, while the information loss caused by the viewpoint change is compensated by enhancing the sorting priority of the upper limb chain features. This dynamic adjustment process is implemented through a programmable logic controller, which can quickly complete the real-time update of the weight coefficients, ensuring the rapid adaptability of the identity recognition features to changes in environmental parameters.

[0083] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0084] The dynamic weight allocation process includes the following steps:

[0085] First, the illumination intensity threshold is set to 100 lux. When the ambient illumination intensity is below 100 lux, the weight of the skeletal point motion trajectory generated by the depth sensor in the local difference feature vector is increased. Specifically, the weight coefficient of the skeletal point motion trajectory feature generated by the depth sensor is increased from the default value, while the weight coefficient of the RGB image feature is correspondingly decreased.

[0086] Secondly, a threshold of 30° is set for the acquisition viewpoint angle. When the acquisition viewpoint angle deviates from the normal viewpoint by more than 30°, the priority of upper limb chain motion features in the global differential feature vector is increased. Specifically, the weight coefficient of upper limb chain motion features is increased from the default value, while the weight coefficient of lower limb chain motion features is correspondingly decreased.

[0087] Through the above technical solution, this application can adaptively adjust feature weights according to actual environmental conditions, relying more on depth information in low-light environments and utilizing upper limb features more in non-frontal viewpoints, thereby improving the robustness and accuracy of the gait recognition system in complex environments. This effectively overcomes the problem of unstable recognition performance in different environments caused by traditional fixed-weight methods, achieving environmental adaptability in gait recognition.

[0088] In some of the solutions described above in this application, when environmental parameters such as light intensity or acquisition viewing angle change significantly, traditional single matching strategies struggle to maintain both recognition accuracy and computational efficiency. Specifically, under low light or side-view conditions, global gait features are susceptible to noise interference, while local limb features may suffer from missing or distorted data, leading to reduced reliability of full feature matching.

[0089] This application further proposes a matching strategy based on environmental parameters, enabling three-level full feature matching under frontal view and strong light conditions; and enabling lower limb feature priority matching under side view or weak light conditions.

[0090] Among them, the three-level full-feature matching achieves progressive feature comparison through a multi-stage screening mechanism, while the lower limb feature-priority matching improves the utilization rate of key features through feature dimension screening. Both are dynamically associated with environmental parameter encoding vectors, and strategy switching is triggered by preset threshold ranges. The selection logic of the matching strategy and the sampling density adjustment in the spatiotemporal segmentation process form a data-level collaboration to ensure the best fit between input features and the matching algorithm.

[0091] Specifically, when environmental parameters meet the requirements of a frontal view and light intensity reaching a strong light level, the system performs a three-level full-feature matching. The first level of matching uses cosine similarity calculation to filter candidate feature subsets from the gait feature library, initially narrowing the comparison range. The second level uses a dynamic time warping algorithm to align and compensate the gait cycle phase of candidate features, eliminating trajectory offset errors caused by differences in step frequency. The third level of matching generates a comprehensive matching index by weighted fusion of cosine similarity and trajectory overlap, ultimately selecting the feature corresponding to the highest value as the judgment result. When environmental parameters enter the side view or weak light range, the system switches to a lower limb feature-priority matching mode. In this mode, gait cycle length and single-step horizontal displacement are first extracted as global stability indicators for initial screening. Subsequently, cosine similarity calculation is performed on the lower limb chain-related feature components, prioritizing the use of lower limb movement features less affected by environmental interference to achieve accurate matching. Through the dynamic selection mechanism of the matching strategy, the system can adaptively adjust the feature comparison dimensions and computational resource allocation in complex environments, effectively balancing recognition accuracy and real-time requirements.

[0092] As a preferred embodiment, the solution of this application is implemented as follows: When the acquisition viewing angle is within the positive viewing angle range and the light intensity is at the strong light level, the three-level full feature matching is activated. In the first-level matching stage, the system calculates the cosine similarity between the identity discrimination features and all samples in the feature library, and selects the top 3% similarity samples to form a candidate subset; in the second-level matching stage, the candidate subset is subjected to dynamic time warping, and the gait cycle phase is compared through a sliding window to calculate the trajectory overlap between the lower limb swing angle and the trunk tilt angle; in the third-level matching stage, the cosine similarity and trajectory overlap are fused according to a preset weight ratio, and the sample with the highest comprehensive score is selected as the judgment result. When the acquisition angle is within the side view range or the light intensity is at a low level, lower limb feature priority matching is enabled: First, the lower limb chain joint motion frequency component and gait cycle length parameter are extracted from the identity discrimination features; preliminary screening is performed by calculating the absolute difference between the gait cycle length and the feature library samples, retaining samples with a difference less than S1 seconds; then, only the lower limb chain joint motion frequency component is cosine similarity calculated, and the sample with the highest similarity is selected as the final judgment result.

[0093] By employing the aforementioned technical solution, this application addresses the problem of decreased recognition accuracy caused by the reliance on a single matching strategy in existing methods under complex environments. By dynamically selecting a feature matching strategy, it fully utilizes complete feature information to improve recognition accuracy under strong lighting conditions, while focusing on robust matching of lower limb features under weak lighting or side-view conditions, effectively reducing the impact of environmental interference on the recognition results. This solution significantly improves the stability of gait recognition under different lighting conditions and acquisition perspectives while maintaining the algorithm's generalization ability.

[0094] In some of the schemes described above in this application, the dynamic weight allocation process adjusts the feature weights according to environmental parameters. However, in actual matching, full feature matching under different environmental conditions may introduce noise interference, leading to decreased matching efficiency and increased misjudgment rate. For example, in low light or side-view conditions, upper limb feature extraction may contain more errors, and using full feature matching would reduce the reliability of recognition.

[0095] This application further proposes a matching strategy based on environmental parameters, enabling three-level full feature matching under frontal view and strong light conditions, and enabling lower limb feature priority matching under side view or weak light conditions.

[0096] The three-level full-feature matching process involves a progressively filtered similarity calculation. First, a candidate subset is established using cosine similarity. Then, dynamic time warping improves phase alignment accuracy. Finally, multi-dimensional similarity is fused to obtain the final result. Lower limb feature-priority matching is implemented in two stages: the first stage uses gait cycle parameters for coarse screening, and the second stage focuses on lower limb features for precise comparison. The correspondence between environmental parameters and matching strategies is determined by preset conditions. When the light intensity is detected to be below a preset threshold or the viewing angle deviates from the orthogonal viewpoint by more than a preset range, the matching strategy is automatically switched.

[0097] Specifically, under normal viewing angle and sufficient lighting, the system executes a three-level full-feature matching process. The first stage calculates the cosine similarity between the feature to be identified and all samples in the database, filtering out candidate sets with similarities exceeding a set threshold. The second stage performs dynamic time warping on the samples within the candidate set to eliminate the impact of gait cycle phase differences on trajectory overlap calculation. The third stage fuses the results from the first two stages according to preset weights to generate a comprehensive matching score ranking. Under side viewing angle or low lighting conditions, the system prioritizes extracting lower limb chain motion features and gait cycle parameters. The first stage performs rapid filtering using gait cycle length and single-step displacement parameters, while the second stage only calculates the cosine similarity of the lower limb motion features of candidate samples. This strategy effectively avoids error interference from upper limb features under adverse conditions; for example, in low lighting, depth sensor data may contain positioning deviations of upper limb joints. Prioritizing matching lower limb features improves recognition reliability.

[0098] As a preferred embodiment, the solution of this application is implemented as follows: In a frontal view scene with an illumination intensity of 1200 lux, a three-level full-feature matching is triggered. In the first-level matching stage, several samples in the pre-stored gait feature library are compared with the identity discrimination features using cosine similarity calculation, and candidate samples with similarity higher than the threshold L1 are selected to form a subset. In the second-level matching stage, the gait cycle sequence of the candidate samples is dynamically time-warped, and the overlap of the skeletal point trajectory on the time axis is calculated using the foot contact point in the gait phase as the alignment reference, where the trajectory overlap threshold is set to L2. In the third-level matching stage, the cosine similarity and trajectory overlap are weighted and summed according to a preset weight ratio, and the sample with the highest comprehensive matching degree is finally selected as the judgment result. Under low-light conditions, when the ambient light intensity drops to 50 lux and the acquisition angle is 45° side view, the lower limb feature priority matching strategy is activated. First, samples with a mean gait cycle length of S2 seconds and a single-step horizontal displacement within the range of [M1, M2] meters are extracted from the pre-stored feature library to filter out candidate samples. Then, the lower limb chain joint angle change sequence of the candidate samples is extracted and cosine similarity is calculated with the lower limb features in the current identity discrimination features. Finally, the sample with the highest similarity is selected to complete the identity determination.

[0099] Through the above technical solution, this application effectively solves the problem of insufficient gait feature matching accuracy in complex environments. Under low light or side-view conditions, prioritizing the matching of lower limb movement features can avoid local feature distortion caused by lighting interference, and improve recognition reliability by utilizing the relatively stable characteristics of human lower limb movement. Under favorable environmental conditions, the multi-level feature fusion matching mechanism significantly reduces the probability of mismatch between similar gaits through spatiotemporal dual verification. This method dynamically adjusts the matching strategy according to different environments, ensuring both the robustness of the recognition system and the optimized allocation of computational resources.

[0100] Example 2:

[0101] like Figure 3 As shown, the gait recognition system based on local and global difference modeling networks includes:

[0102] The data acquisition module is used to acquire the gait contour sequence and environmental parameters of the target object;

[0103] The data processing module is used to adjust the segmentation granularity according to the environmental parameters, perform spatiotemporal segmentation processing on the gait contour sequence, and generate local limb movement trajectory fragments and global gait cycle sequences.

[0104] The feature extraction module is used to perform frequency domain feature transformation on the local limb movement trajectory segment to obtain a local difference feature vector, and to perform spatial alignment encoding on the global gait cycle sequence to obtain a global difference feature vector;

[0105] The feature fusion module is used to perform cross-scale interactive calculations on the local difference feature vector and the global difference feature vector to generate fused difference features with spatiotemporal consistency; and to dynamically assign weights to the fused difference features according to the environmental parameters to obtain optimized identity discrimination features.

[0106] The strategy selection module is used to select a matching strategy based on the environmental parameters; the matching strategy includes three-level full feature matching and lower limb feature-priority matching.

[0107] The feature matching module is used to perform similarity matching between the identity discrimination features and the pre-stored gait feature library according to the corresponding matching strategy, and output the identity determination result of the target object.

[0108] The above description is merely an example and illustration of the structure of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the structure of the invention or exceed the scope defined in the claims, all of which should fall within the protection scope of the present invention.

[0109] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0110] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A gait recognition method based on local and global difference modeling networks, characterized in that, Includes the following steps: The system acquires the gait contour sequence and environmental parameters of the target object. The environmental parameters include light intensity and acquisition viewing angle. The light intensity is measured in real time using a photosensitive sensor built into a visible light camera. The angle between the centroid of the target object and the normal of the acquisition plane is calculated using a depth sensor and used as the acquisition viewing angle. The light intensity is divided into three levels: strong light, normal light, and weak light, according to a preset light intensity threshold. The acquisition viewing angle is divided into two azimuth intervals: frontal view and side view, according to a preset angle threshold. Establish a timestamp synchronization mechanism between environmental parameters and gait profile sequences to ensure the timeliness consistency of parameter association; The segmentation granularity is adjusted according to the environmental parameters, and the gait contour sequence is subjected to spatiotemporal segmentation to generate local limb movement trajectory fragments and global gait cycle sequences. The spatiotemporal segmentation process includes: dividing the gait contour sequence into N phase intervals, each phase interval corresponding to the spatiotemporal segmentation parameters of the foot touch-down, leg swing, and take-off phases; within a single phase interval, dividing the human body into three local regions—upper limb chain, trunk chain, and lower limb chain—using a kinematic chain model; and when sampling the motion trajectory for each local region, using a time window sliding mechanism to ensure that the overlap rate of adjacent trajectory segments is not lower than a preset threshold. The process of adjusting the segmentation granularity based on environmental parameters includes: using a fine-grained segmentation mode under strong light to increase the sampling density of upper limb chain motion trajectories; using a coarse-grained segmentation mode under weak light to retain the sampling data of lower limb chain motion trajectories; and retaining the sampling data of lower limb chain motion trajectories from a side view. The local limb movement trajectory segment is subjected to frequency domain feature transformation to obtain a local difference feature vector, and the global gait cycle sequence is subjected to spatial alignment encoding to obtain a global difference feature vector; The local difference feature vector and the global difference feature vector are subjected to cross-scale interactive calculation to generate a fused difference feature with spatiotemporal consistency. The cross-scale interactive computation includes establishing a dynamic mapping relationship between local differential feature vectors and global differential feature vectors, and calculating the contribution weight of each local region to the global features through a spatial attention mechanism. The cross-scale interactive computing also includes an environment-aware attention mechanism: gating coefficients are generated based on the environment parameter encoding vector E, and the calculation process satisfies: Aenv=σ(Wl*L+Wg*G+We*E), where L is the local differential feature vector, G is the global differential feature vector, Wl, Wg, and We are trainable weight matrices, σ is the sigmoid activation function, and Aenv is used as a dynamic weight to control the fusion ratio of local and global features. The method for generating the environmental parameter encoding vector E includes: performing a logarithmic transformation on the illumination intensity to obtain a normalized illumination encoding value; performing sine and cosine encoding on the acquisition viewing angle to generate an azimuth feature vector; and concatenating the two types of encoding vectors and then reducing their dimensionality through a fully connected layer to form the environmental parameter encoding vector E. Based on the environmental parameters, the fusion difference features are dynamically weighted to obtain optimized identity discrimination features; A matching strategy is selected based on the environmental parameters; the matching strategy includes three-level full feature matching and lower limb feature-priority matching. According to the corresponding matching strategy, the identity discrimination features are matched with the pre-stored gait feature library for similarity, and the identity determination result of the target object is output.

2. The gait recognition method based on local and global difference modeling networks according to claim 1, characterized in that: The process of obtaining gait contour sequences includes: Gait contour sequences are acquired by a multi-source sensing device, wherein the multi-source sensing device includes at least a visible light camera and a depth sensor; The system acquires RGB gait image sequences using a visible light camera, while simultaneously using a depth sensor to obtain the motion trajectory of three-dimensional skeletal points. When extracting contours from the RGB gait image sequence, an adaptive illumination compensation algorithm is used to eliminate pixel distortion in low-light areas; The motion trajectory of the three-dimensional skeletal points is kinematically constrained and verified to remove abnormal data points that exceed the human joint activity threshold.

3. The gait recognition method based on local and global difference modeling networks according to claim 2, characterized in that: The frequency domain feature transformation process includes: Short-time Fourier transform is performed on local limb movement trajectory segments to extract the motion frequency distribution features of each joint. The motion frequency distribution features are reduced in dimensionality using principal component analysis, retaining feature components whose variance contribution rate exceeds a preset threshold. The global gait cycle sequence is divided into a three-dimensional spatial grid, and the density gradient change characteristics of bone points in each grid cell are calculated. The calculation method for the density gradient change characteristics includes: The three-dimensional spatial mesh is decomposed into horizontal slice units and vertical slice units along the direction of motion; Calculate the spatial distribution dispersion of bone points within each slice unit to generate a density gradient change curve; Differential calculations are performed on the density gradient change curves of adjacent slice units, and curvature abrupt change points are extracted as feature markers.

4. The gait recognition method based on local and global difference modeling networks according to claim 3, characterized in that: The dynamic weight allocation process includes: When the light intensity is lower than a preset threshold, increase the weight of the motion trajectory of the skeleton points generated by the depth sensor in the local difference feature vector; When the acquisition viewpoint deviates from the orthogonal viewpoint by more than a preset range, the priority of the upper limb chain motion features in the global differential feature vector is enhanced.

5. The gait recognition method based on local and global difference modeling networks according to claim 4, characterized in that: The process of selecting a matching strategy based on environmental parameters includes: Enables three-level full feature matching under positive viewing angle and strong lighting conditions; Lower limb features are prioritized for matching in side view or low light conditions.

6. The gait recognition method based on local and global difference modeling networks according to claim 5, characterized in that: The three-level full feature matching includes: First-level matching: Calculate the cosine similarity between the identity discrimination features and all samples in the pre-stored gait feature library, and filter out the candidate feature subset; Second-level matching: Dynamic time warping is performed on the candidate feature subset to calculate the trajectory overlap after gait cycle phase alignment; Third-level matching: The cosine similarity and trajectory overlap are fused to generate a comprehensive matching score, and the feature with the highest comprehensive matching score is selected as the judgment result; The process of prioritizing lower limb feature matching includes: Lower limb chain-related feature components are extracted from local differential feature vectors, and gait cycle length and single-step horizontal displacement are extracted from global differential feature vectors. First-level matching: Calculate the similarity between the gait cycle length and single-step horizontal displacement and all samples in the pre-stored gait feature library, and filter out candidate feature subsets; Second-level matching: The cosine similarity of the candidate feature subset with the limb chain-related feature components is calculated, and the feature with the highest cosine similarity is used as the judgment result.

7. A gait recognition system based on local and global difference modeling networks, characterized in that: Using a gait recognition method based on local and global difference modeling networks as described in any one of claims 1 to 6, comprising: The data acquisition module is used to acquire the gait contour sequence and environmental parameters of the target object; The data processing module is used to adjust the segmentation granularity according to the environmental parameters, perform spatiotemporal segmentation processing on the gait contour sequence, and generate local limb movement trajectory fragments and global gait cycle sequences. The feature extraction module is used to perform frequency domain feature transformation on the local limb movement trajectory segment to obtain a local difference feature vector, and to perform spatial alignment encoding on the global gait cycle sequence to obtain a global difference feature vector; The feature fusion module is used to perform cross-scale interactive calculations on the local difference feature vector and the global difference feature vector to generate fused difference features with spatiotemporal consistency; and to dynamically assign weights to the fused difference features according to the environmental parameters to obtain optimized identity discrimination features. The strategy selection module is used to select a matching strategy based on the environmental parameters; the matching strategy includes three-level full feature matching and lower limb feature-priority matching. The feature matching module is used to perform similarity matching between the identity discrimination features and the pre-stored gait feature library according to the corresponding matching strategy, and output the identity determination result of the target object.

Images