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

Abstract

The invention discloses a gait recognition method and system based on a local and global difference modeling network, and relates to the field of gait recognition, and the method comprises the steps: obtaining a gait contour sequence and environment parameters of a target object; processing the gait contour sequence according to environmental parameters to generate local limb movement track fragments and a global gait periodic sequence; carrying out feature extraction on the local limb movement track fragment and the global gait periodic 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 features according to the environmental parameters to obtain identity discrimination features; selecting a matching strategy according to the environmental parameters, performing similarity matching on the identity judgment features and a pre-stored gait feature library according to the matching strategy, and outputting an identity judgment result of the target object; the method can overcome the interference of a complex environment, and improves the stability of feature extraction.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of gait recognition, in particular to a gait recognition method and system based on local and global difference modeling network. BACKGROUND

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

[0003] In actual application scenarios, complex environmental factors such as light changes and view angle changes will significantly affect the extraction and recognition effect of gait features. Traditional methods perform poorly when environmental parameters change, especially in low light conditions, the quality of gait images obtained by image acquisition devices decreases, leading to inaccurate feature extraction; when pedestrians walk at different viewing angles, the appearance features of gait will change greatly, making the model trained based on fixed viewing angles have a large decrease in recognition accuracy. These problems seriously restrict the application effect of gait recognition technology in real scenes.

[0004] Existing methods have limitations in feature extraction, many methods focus on extracting global gait features, ignoring the importance of local detail features. Although global features can reflect the overall pattern of gait, they are insufficient to capture some subtle, individual-specific local features. These local features play a key role in distinguishing similar gaits, and lack of effective extraction of them will limit the accuracy of gait recognition. At the same time, existing methods lack effective feature fusion mechanisms, making it difficult to achieve collaborative optimization of local features and global features.

[0005] In view of the above problems, the existing technology needs to be improved SUMMARY

[0006] In view of the shortcomings of the prior art, the present application provides a gait recognition method and system based on local and global difference modeling network.

[0007] In order to achieve the above purpose, the technical scheme of the present application is as follows:

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

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

[0010] Adjust the segmentation granularity according to the environmental parameters, perform spatiotemporal segmentation processing on the gait contour sequence, and generate local limb motion trajectory segments and global gait cycle sequences;

[0011] performing frequency domain feature transformation on the local limb motion trajectory segment to obtain a local difference feature vector, and performing spatial alignment coding on the global gait cycle sequence to obtain a global difference feature vector;

[0012] performing cross-scale interaction calculation on the local difference feature vector and the global difference feature vector to generate a fusion difference feature with spatio-temporal consistency;

[0013] performing dynamic weight distribution on the fusion difference feature according to the environment parameter to obtain an optimized identity discrimination feature;

[0014] selecting a matching strategy according to the environment parameter; the matching strategy includes three-level full feature matching and lower limb feature priority matching;

[0015] performing similarity matching on the identity discrimination feature and a pre-stored gait feature library according to the corresponding matching strategy, and outputting an identity determination result of the target object.

[0016] In a second aspect, the present application discloses a gait recognition system based on local and global difference modeling network, comprising:

[0017] a data acquisition module configured to acquire a gait contour sequence of a target object and an environment parameter;

[0018] a data processing module configured to adjust a segmentation granularity according to the environment parameter, perform spatio-temporal segmentation processing on the gait contour sequence, and generate a local limb motion trajectory segment and a global gait cycle sequence;

[0019] a feature extraction module configured to perform frequency domain feature transformation on the local limb motion trajectory segment to obtain a local difference feature vector, and perform spatial alignment coding on the global gait cycle sequence to obtain a global difference feature vector;

[0020] a feature fusion module configured to perform cross-scale interaction calculation on the local difference feature vector and the global difference feature vector to generate a fusion difference feature with spatio-temporal consistency, and perform dynamic weight distribution on the fusion difference feature according to the environment parameter to obtain an optimized identity discrimination feature;

[0021] a strategy selection module configured to select a matching strategy according to the environment parameter; the matching strategy includes three-level full feature matching and lower limb feature priority matching;

[0022] a feature matching module configured to perform similarity matching on the identity discrimination feature and a pre-stored gait feature library according to the corresponding matching strategy, and output an identity determination result of the target object.

[0023] Compared with the prior art, the present application has the following advantages:

[0024] 1. By real-time monitoring of light intensity and viewing angle, dynamically adjusting the spatiotemporal segmentation granularity (such as increasing the upper limb sampling density under strong light, focusing on lower limb features under weak light / side view), effectively overcoming complex environmental interference, and improving the stability of feature extraction;

[0025] 2. Fusion of visible light camera (RGB image) and depth sensor (three-dimensional skeleton point) data, under low light conditions, preferentially using depth data to compensate for contour information, combining kinematic constraints to check and remove abnormal points, ensuring the integrity and rationality of the input data;

[0026] 3. Using short-time Fourier transform to extract local joint frequency domain features, combining global skeleton point density gradient analysis, and through environmental perception attention mechanism to fuse local and global features, enhancing the spatiotemporal consistency expression ability;

[0027] 4. According to the environmental parameters, automatically switch the matching mode: strong light / normal view angle enable three-level full feature matching, weak light / side view angle enable lower limb priority matching, suppress the features of upper limbs, trunk, etc. which are easily affected by viewing angle / shading, when the viewing angle deviates or the light is insufficient, use the anti-interference characteristics of lower limb movement to improve the recognition robustness. BRIEF DESCRIPTION OF DRAWINGS

[0028] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description only constitute some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0029] Figure 1 The overall method block diagram of the first embodiment of the present application;

[0030] Figure 2 The method flowchart of the first embodiment of the present application;

[0031] Figure 3 The overall system block diagram of the second embodiment of the present application. DETAILED DESCRIPTION

[0032] The technical solutions of the present application will be described below in conjunction with the embodiments. Obviously, the described embodiments are only some of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0033] Summary of the application: In the traditional existing gait recognition method, the change of environmental parameters causes deviation in the feature extraction process, the spatio-temporal correlation of global and local features is insufficient, and the multi-scale feature fusion in dynamic scenes lacks adaptability. Under complex lighting conditions, the contour extraction algorithm cannot eliminate pixel distortion in low light areas, the spatial distribution of skeletal points is abnormal due to perspective shift, the time-space segmentation process does not establish a correlation mechanism with physical environmental parameters, and the cross-scale feature interaction process does not introduce an environmental perception mechanism, resulting in representation errors when dynamic weight allocation is performed on the fused difference features.

[0034] For example, in a gait recognition device deployed in a security monitoring system, data is synchronously collected by a visible light camera and a depth sensor. When the ambient light intensity decreases and the collection perspective deviates from the normal direction by a certain angle, the gait sequence output by the contour extraction module has broken lower limb joint trajectory and misaligned upper limb movement phase markers. The fixed granularity parameter used by the time-space segmentation processing module causes over-sampling of the upper limb chain movement trajectory in strong light environments, resulting in noise interference, and under-sampling of the lower limb chain movement trajectory in low light environments, resulting in 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 do not match in scale, and the mapping relationship between environmental parameters and feature weights is not established during the cross-scale interaction process. The final identity discrimination feature has a cosine similarity calculation deviation when it is dynamically matched.

[0035] If the above problems are not solved, the identity determination module will not be able to accurately distinguish individual differences in similar gait patterns, resulting in an increase in misidentification rate in cross-environment scenarios. The static features stored in the gait feature library cannot adapt to feature drift in dynamic environments, the timestamps of multi-source sensor data are asynchronous, causing feature fusion phase misalignment, and the robustness of the system in complex scenarios is significantly reduced. Feature representation mismatch caused by perspective changes during the identity authentication process will cause security vulnerabilities, and the real-time discrimination ability of abnormal gait behavior in monitoring scenarios is constrained.

[0036] In the face of the above problems, the present application first considers establishing a dynamic correlation mechanism between environmental parameters and data processing. To solve the problem of feature misalignment caused by changes in light and perspective, the sensor data is time-stamped synchronously with the environmental parameters, and the feature extraction strategy is dynamically adjusted. To solve the problem of insufficient spatio-temporal correlation between global and local features, a cross-scale interaction mechanism is introduced after frequency domain transformation and spatial encoding, so that local movement trajectories and global periodic sequences form complementary features. For the problem of dynamic weight allocation deviation, a gated fusion method based on environmental parameter encoding vectors is designed, which controls the feature fusion ratio through a trainable weight matrix. Finally, the time-space segmentation granularity is linked with the environmental parameters to establish a segmentation mode adaptive adjustment mechanism, and a complete technical link is formed by combining cross-scale feature interaction and dynamic weight allocation.

[0037] Embodiment one:

[0038] like Figures 1-2 As shown, the gait recognition method based on the local and global difference modeling network includes the following steps: obtaining the gait contour sequence and environmental parameters of the target object; adjusting the segmentation granularity according to the environmental parameters, performing spatiotemporal segmentation processing on the gait contour sequence, and generating local limb motion trajectory segments and global gait cycle sequences; performing frequency domain feature transformation on the local limb motion 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 the 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 on the identity discrimination features and the pre-stored gait feature library according to the corresponding matching strategy, and outputting the identity determination result of the target object.

[0039] The present application further proposes obtaining a gait contour sequence through a multi-source sensing device, which includes at least a visible light camera and a depth sensor; collecting an RGB gait image sequence through the visible light camera, while simultaneously obtaining a three-dimensional skeletal point motion trajectory through the depth sensor; using an adaptive illumination compensation algorithm to eliminate pixel distortion in low-light areas when performing contour extraction on the RGB gait image sequence; and performing kinematic constraint verification on the three-dimensional skeletal point motion trajectory to eliminate abnormal data points that exceed the threshold of human joint movement.

[0040] The multi-source sensing device's visible light camera and depth sensor utilize a synchronized triggering mechanism to ensure alignment of RGB image timestamps with the skeletal point trajectory. An adaptive illumination compensation algorithm dynamically adjusts the gamma correction coefficient based on the image histogram distribution, enhancing local brightness in low-light areas. Kinematic constraint verification establishes a joint angle change rate model to detect sudden changes in skeletal point trajectories that deviate from human motion patterns.

[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. After calculating the average brightness of each block, a correction curve with a preset gamma value is applied to areas below 50 lux to enhance dark details. The depth sensor outputs the three-dimensional 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 complementation 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 are used to ensure the physiological rationality of the motion trajectory, providing high-quality input for subsequent spatiotemporal segmentation.

[0042] Through the technical solutions, the gait image under a complex lighting environment is effectively processed, and the accuracy of contour extraction is improved. Meanwhile, the effectiveness of the skeleton point data is ensured through kinematic constraint verification, laying a foundation for subsequent feature extraction. Therefore, the robustness and adaptability of the gait recognition system under different environmental conditions are improved.

[0043] In some solutions of the present application, the acquisition method of the environmental parameters has the problems of insufficient timeliness and inaccurate parameter division, which leads to the inability to accurately associate the influencing factors of environmental changes in the gait feature extraction process. For example, the non-graded division of the light intensity may make the subsequent algorithm unable to effectively distinguish the feature processing mode under different lighting conditions, the non-azimuth interval classification of the viewing angle may lead to the decline of the accuracy of the spatial alignment coding, and the asynchronous parameters and time stamps of the gait data may cause the timing dislocation in the data processing.

[0044] The present application further proposes that the acquisition process of the environmental parameters includes light intensity and collection viewing angle; the light intensity is measured in real time through the photosensitive sensor built in the visible light camera; the normal angle between the mass center of the target object and the collection plane is calculated through the depth sensor, as the collection viewing angle; the light intensity is divided into three graded intervals of strong light, normal light and weak light according to the preset light intensity threshold; the collection viewing angle is divided into two azimuth intervals of front viewing angle and side viewing angle according to the preset angle threshold; and a time stamp synchronization mechanism of the environmental parameters and the gait contour sequence is established to ensure the timeliness consistency of the parameter association.

[0045] The real-time measurement of the light intensity is realized through the photosensitive sensor, the response frequency of which is matched with the frame rate of the camera to ensure the real-time of the light data corresponding to each frame of image. The collection viewing angle is obtained through the geometric calculation of the normal angle between the mass center and the collection plane. Specifically, the normal angle is calculated after the plane equation is constructed through the three-dimensional coordinate data output by the depth sensor. The light intensity threshold division adopts a fixed numerical interval, for example, the strong light interval is greater than 1000 lux, the normal light interval is 300-1000 lux, and the weak light interval is less than 300 lux. The angle threshold adopts a preset 30 degrees as the dividing point of the front viewing angle and the side viewing angle. When the angle is less than 30 degrees, it is determined as the front viewing angle, otherwise as the side viewing angle. The time stamp synchronization mechanism is triggered by the hardware clock to align the time of environmental parameter acquisition and the time of gait contour sequence generation, with an error controlled within milliseconds.

[0046] Specifically, the photosensitive sensor is integrated inside the camera module, directly reading the voltage signal output by the sensor and converting it into a light intensity value, avoiding external environmental interference. The depth sensor calculates the center of mass position through the three-dimensional coordinates of the skeletal points, taking the normal vector of the collection plane as the reference, and uses the vector dot product formula to calculate the cosine value of the included angle. The light intensity level division is achieved by comparing the real-time measurement value with the preset threshold interval, for example, when the measurement value is 500 lux, it is determined as normal light level. The view angle orientation interval division is achieved by comparing the included angle calculation result with the 30-degree threshold, for example, when the included angle is 45 degrees, it is determined as a side view angle. The timestamp synchronization is achieved through a synchronous signal triggering mechanism, when the camera and depth sensor collect data, the central processor records a unified system timestamp, ensuring that the timing of the environmental parameters and gait data strictly corresponds in subsequent processing. Thus, the quantification classification of light intensity and view angle provides accurate parameter input for subsequent dynamic adjustment of segmentation granularity and matching strategy, and the timestamp synchronization mechanism eliminates the delay error of data correlation, ensuring the spatiotemporal consistency of feature extraction.

[0047] As a preferred embodiment, the scheme of the present application is implemented as follows:

[0048] The environmental parameters include light intensity and collection view angle. The light intensity is measured in real time by the photosensitive sensor built-in the visible light camera. The collection view angle is obtained by calculating the normal included angle between the target object center of mass and the collection plane through the depth sensor. The light intensity is divided into three level intervals according to the preset light intensity threshold: strong light, normal light, and weak light. The collection view angle is divided into two orientation intervals according to the preset angle threshold: front view angle and side view angle. A timestamp synchronization mechanism is established between the environmental parameters and the gait contour sequence to ensure the time consistency of parameter correlation.

[0049] Specifically, the photosensitive sensor built-in the visible light camera collects light intensity data every 100 milliseconds. The depth sensor calculates the normal included angle between the target object center of mass and the collection plane every 50 milliseconds. The light intensity is divided into three level intervals: 0-100 lux for weak light, 100-1000 lux for normal light, and 1000 lux or above for strong light. The collection view angle is divided into two orientation intervals: 0-30 degrees for front view angle and 30-90 degrees for side view angle. The timestamp synchronization mechanism between the environmental parameters and the gait contour sequence uses a timestamp alignment method to match the collection timestamp of the environmental parameters with the nearest gait contour sequence frame timestamp.

[0050] By the technical solution, the environment parameter is accurately acquired and quantified. Therefore, the gait feature extraction and matching strategy can be dynamically adjusted according to the real-time illumination and viewing angle conditions, and the adaptability of the gait recognition system to complex environments is improved. Further, by establishing a time synchronization mechanism between the environment parameter and the gait data, the consistency of the environment factors and the gait data in the feature extraction process is ensured, and the feature distortion caused by the lag of the environment parameter is avoided.

[0051] In some schemes of the application, the environment parameter is divided into different level intervals, but it is not clear how to adjust the granularity of space-time segmentation based on different environment parameters, resulting in difficulty in balancing the contradiction between the integrity of local feature extraction and data redundancy in complex environment conditions. For example, when the gait contour sequence collected under low light conditions has noise interference, if a fixed segmentation mode is still used, it may lead to too high sampling density of the local limb motion trajectory, introducing invalid feature components.

[0052] The application further divides the gait contour sequence into N phase intervals, each phase interval corresponding to the space-time segmentation parameters of the foot landing, leg swinging and emptying phases; in a single phase interval, the human body is divided into three local regions of upper limb chain, torso chain and lower limb chain through a kinematic chain model; when sampling the motion trajectory of each local region, a time window sliding mechanism is used to ensure that the overlap rate of adjacent trajectory segments is not less than a preset threshold; the process of adjusting the segmentation granularity according to the environment parameter 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; retaining the sampling data of the lower limb chain motion trajectory when the viewing angle is side.

[0053] The division of the phase interval is based on the biomechanical characteristics in the human gait cycle, and each phase interval corresponds to a specific kinematic parameter threshold. The kinematic chain model abstracts the human body as a multi-rigid-body system, and establishes the motion correlation between the chain segments through joint constraint equations. The time window sliding mechanism uses a variable window length, and the window moving step is dynamically adjusted according to the gait frequency, and the preset overlap rate threshold is set to P1%. The segmentation granularity adjustment module has a strategy mapping table under different combinations of environment parameters, for example, when the light intensity level is weak light and the viewing angle is side, the lower limb chain trajectory retention mode is automatically triggered.

[0054] Specifically, under strong light conditions, the quality of the RGB image collected by the visible light camera is high, and the fine-grained segmentation mode improves the sampling frequency of the shoulder joint swing trajectory by shortening the upper limb chain sampling interval, for example, adjusting the sampling interval of the upper limb chain from 10 ms per frame to 5 ms. Under weak light conditions, the skeletal point data collected by the depth sensor is less affected by noise, and the coarse-grained segmentation mode adjusts the sampling interval of the lower limb chain to 20 ms, while discarding the trajectory points in the upper limb chain with a confidence value below the preset confidence threshold. When the viewing angle is a side viewing angle, the projection of the motion trajectory of the upper limb chain in the three-dimensional coordinate system has a large deformation, and at this time, the lower limb chain data is retained and the knee joint angle change curve is reconstructed by a kinematics inverse solution algorithm. The time window sliding mechanism generates an overlapping area between adjacent trajectory segments, and a Kalman filter is used to smooth the trajectory data in the overlapping area, eliminating trajectory discontinuities caused by sensor jitter.

[0055] As a preferred embodiment, the scheme of the present application is implemented as follows:

[0056] The gait profile sequence is divided into 8 phase intervals, each corresponding to the spatial and temporal segmentation parameters of the foot touch, leg swing, and empty phase. Within a single phase interval, the human body is divided into three local regions: the upper limb chain, the torso chain, and the lower limb chain, through a kinematics chain model. When sampling the motion trajectory of each local region, the time window sliding mechanism is used to ensure that the overlap rate of adjacent trajectory segments is not less than 80%.

[0057] The process of adjusting the segmentation granularity according to environmental parameters includes: under strong light conditions with an illumination intensity greater than 1000 lux, using a fine-grained segmentation mode to increase the sampling density of the upper limb chain motion trajectory to 60 frames per second; under weak light conditions with an illumination intensity less than 100 lux, using a coarse-grained segmentation mode to retain only the sampling data of the lower limb chain motion trajectory, and reducing the sampling density to 30 frames per second; when the collection viewing angle is greater than 45 degrees from the front direction, also retaining only the sampling data of the lower limb chain motion trajectory.

[0058] Through the above technical scheme, the present application can adaptively adjust the granularity of gait feature extraction according to different environmental conditions. Under strong light conditions, detailed information is fully utilized to improve recognition accuracy; under weak light or side viewing angle conditions, the sampling granularity is reduced and the lower limb features are focused 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 variable environments.

[0059] In some of the above schemes of the present application, when feature extraction is performed based on the local limb motion trajectory segments and the global gait cycle sequence after spatiotemporal segmentation processing, it is difficult to effectively capture the dynamic changes of the local joint motion frequency and accurately represent the key gradient changes of the skeletal point spatial distribution, resulting in insufficient feature discrimination.

[0060] The application further proposes to perform short-time Fourier transform on the local limb motion trajectory segment, extract the motion frequency distribution characteristics of each joint point, perform dimension reduction processing on the motion frequency distribution characteristics through principal component analysis, and retain the characteristic components with a variance contribution rate exceeding a preset threshold. The global gait cycle sequence is divided into a three-dimensional space grid, and the density gradient change characteristics of the skeletal points in each grid unit are calculated. The calculation method of the density gradient change characteristics includes: decomposing the three-dimensional space grid into a transverse slice unit and a longitudinal slice unit along the motion direction; calculating the spatial distribution dispersion of the skeletal points in each slice unit to generate a density gradient change curve; and performing difference calculation on the density gradient change curves of adjacent slice units to extract the curvature mutation points as feature markers.

[0061] In the short-time Fourier transform, a fixed time window is used to perform frequency spectrum analysis on the time domain signal of the local trajectory segment, the window length is set to one quarter of the gait cycle length, and the overlap rate is not less than P1%. In the principal component analysis dimension reduction, the preset variance contribution rate threshold is P2%, and the first three principal component components are retained. In the three-dimensional space grid division, the space is divided into cubic units with a size of 20 cm x 20 cm x 20 cm along the human motion direction, the transverse slice unit is cut along the sagittal plane, and the longitudinal slice unit is cut along the coronal plane. The spatial distribution dispersion is obtained by calculating the standard deviation of the skeletal point coordinates in the slice unit, and the density gradient change curve is arranged in time sequence. The difference calculation uses a first-order forward difference operator, and the curvature mutation point is defined as the position where the difference value exceeds twice the standard deviation of the mean value.

[0062] Specifically, after the local limb motion trajectory segment is subjected to short-time Fourier transform, the motion frequency distribution characteristics of the joint points are decomposed into amplitude spectrum and phase spectrum, wherein the amplitude spectrum is used to represent the motion intensity, and the phase spectrum is used to represent the motion timing relationship. The principal component analysis method selects the low-frequency components with the strongest representation ability through eigenvalue decomposition of the covariance matrix. The three-dimensional space grid division maps the global gait cycle sequence to a three-dimensional space, the transverse slice unit captures the lateral motion characteristics, and the longitudinal slice unit captures the front and rear motion characteristics. The skeletal point spatial distribution dispersion reflects the concentration degree of the motion trajectory, and the density gradient change curve forms a dynamic pattern through the accumulation in the time dimension. The difference calculation effectively amplifies the distribution difference between adjacent slice units, and the curvature mutation point corresponds to the key action turning moment in the gait cycle. Through the above steps, the local joint motion frequency characteristics and the global skeletal density gradient characteristics are complementary, and the representation ability of the characteristics is enhanced.

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

[0064] The frequency domain feature transformation process includes performing a short-time Fourier transform on the local limb motion trajectory segment, and extracting the motion frequency distribution features of each joint.

[0065] The global gait cycle sequence is divided into a three-dimensional spatial grid, and the density gradient change features of the skeletal points in each grid cell are calculated. The calculation method of the density gradient change features includes decomposing the three-dimensional spatial grid into transverse slice cells and longitudinal slice cells along the motion direction. Specifically, the spatial distribution dispersion of the skeletal points in each slice cell is calculated to generate a density gradient change curve. Thus, the density gradient change curves of adjacent slice cells are calculated by difference, and the curvature mutation points are extracted as feature markers.

[0066] For example, in actual application, the gait cycle sequence can be divided into a 10x10x10 three-dimensional grid. For each grid cell, the spatial distribution dispersion of the skeletal points therein is calculated to form a 100x100 density gradient matrix. Then the matrix is decomposed into 10 10x10 slice cells along the motion direction. The standard deviation of the skeletal point distribution is calculated for each slice cell to obtain a density gradient change curve composed of 10 standard deviation values. Finally, the density gradient change curves of adjacent slice cells are differentiated, and points with a differential value greater than a preset threshold are extracted as curvature mutation points, which are the feature markers.

[0067] Through the above technical solutions, the frequency domain features of local limb motion and the spatial features of global gait cycle can be effectively extracted. The frequency domain feature transformation can capture the periodic patterns of joint motion, while the density gradient change features reflect the spatial distribution rules of the overall gait. This multi-scale feature extraction method improves the expression ability of gait features, enabling the model to focus on both local details and global patterns, thereby enhancing the accuracy and robustness of gait recognition. In particular, in complex environments such as changes in lighting or viewing angle, this method still maintains good feature extraction effect, improving the adaptability of the gait recognition system to different scenes.

[0068] In some of the above schemes of the present application, in the cross-scale interaction calculation process, the influence of environmental changes on the fusion of local and global features is not fully modeled, resulting in a lack of adaptability in feature weight distribution under different environmental conditions, affecting the discriminability of the fused features.

[0069] The application further proposes to establish a dynamic mapping relationship of local difference feature vectors and global difference feature vectors, to calculate the contribution weight of each local region to the global feature through a spatial attention mechanism; the cross-scale interaction calculation further includes an environment perception attention mechanism: a gating coefficient is generated according to an environment parameter encoding vector E, and the calculation process satisfies Aenv=σ(WlL+WgG+We*E), wherein L is a local difference feature vector, G is a global difference feature vector, Wl, Wg, and We are trainable weight matrices, and sigma is a sigmoid activation function, and Aenv is taken as a dynamic weight to control the fusion proportion of local and global features; the generation method of the environment parameter encoding vector E includes: performing logarithmic transformation on the light intensity to obtain a normalized light encoding value; performing sine-cosine encoding on the collected viewing angle to generate an azimuth feature vector; and concatenating the two types of encoding vectors and then reducing the dimension through a fully connected layer to form the environment parameter encoding vector E.

[0070] The dynamic mapping relationship is realized through a spatial attention mechanism, and the correlation weight of each local limb region motion trajectory segment to the global gait cycle sequence is calculated. The environment perception attention mechanism applies a gating coefficient Aenv to the feature fusion process, and performs linear combination on the local feature L, the global feature G, and the environment parameter E through a trainable weight matrix. The logarithmic transformation of the light intensity adopts a logarithm function with a base of 10 to compress the light intensity value to the interval of 0-1; the sine-cosine encoding of the viewing angle converts the angle theta into a two-dimensional vector [sin theta, cos theta]. The fully connected layer dimension reduction operation maps the concatenated encoding vector to a 128-dimensional vector space.

[0071] Specifically, under strong light conditions, the normalized light encoding value tends to 1, at this time, the gating coefficient Aenv is more inclined to retain the upper limb motion details in the local difference feature; under the side viewing angle condition, the azimuth feature vector forms a two-dimensional vector through the sine-cosine encoding, and the environment parameter encoding vector E is formed after the light encoding is concatenated and processed through the fully connected layer, the vector adjusts the gating coefficient Aenv through the weight matrix We, so that the fusion process pays more attention to the lower limb motion trajectory. The dynamic weight Aenv output by the sigmoid activation function is subjected to Hadamard product operation with the feature vector to realize the environment-adaptive feature fusion proportion adjustment.

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

[0073] A dynamic mapping relationship between the local difference feature vector and the global difference feature vector is established, and a spatial attention mechanism is used to calculate the contribution weight of each local region to the global feature. The cross-scale interaction calculation also includes an environment perception attention mechanism. A gating coefficient is generated according to the environment parameter encoding vector E, and the calculation process satisfies: Aenv=σ(WlL+WgG+We*E), L is a local difference feature vector, G is a global difference feature vector, Wl, Wg, and We are trainable weight matrices, and sigma is a 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 logarithmic transformation on the light intensity to obtain a normalized light encoding value; performing sine-cosine encoding on the collection angle to generate an azimuth feature vector; and concatenating the two types of encoding vectors and reducing the dimension through a fully connected layer to form the environment parameter encoding vector E.

[0074] Specifically, first, a spatial attention module is constructed, which includes multiple convolution layers and nonlinear activation functions. The local difference feature vector L is input into the module to obtain the attention weight of each local region. Then, the weight is weighted and summed with the global difference feature vector G to obtain the preliminary fusion feature.

[0075] Further, an environment perception attention module is constructed. The module first performs logarithmic transformation on the light intensity, mapping the light intensity range of 0-10000 lux to 0-1. The collection angle is encoded by sine-cosine, and the angle of 0-360 degrees is encoded into a two-dimensional vector. The light encoding and the angle encoding are concatenated, and then reduced in dimension through a fully connected layer including 64 neurons to obtain 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 as random values. Sigma is a sigmoid activation function that maps the output to 0-1. The obtained Aenv is the dynamic fusion weight of environment perception.

[0077] Finally, the preliminary fusion feature is multiplied by Aenv to obtain the final fusion difference feature. The feature contains local details, global structures, and environmental factors.

[0078] By the technical solution, adaptive fusion of local features and global features is realized, and dynamic adjustment of environmental factors is introduced. The fusion mode can automatically adjust the importance of features according to different environmental conditions, improving the robustness of gait recognition in complex environments. At the same time, the spatial attention mechanism highlights the contribution of key local areas, enhancing the model's ability to capture subtle differences. The environmental perception 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 above schemes of the application, when the light intensity is low or the collection viewing angle deviates from the normal viewing angle, the weight distribution of different data sources in the fusion difference features fails to dynamically adjust according to environmental changes, resulting in insufficient robustness of identity discrimination features in complex environments, affecting the final matching accuracy. For example, under low light conditions, the quality of the gait contour sequence obtained by the visible light camera decreases, at this time the skeletal point motion trajectory generated by the depth sensor has higher credibility, but the existing method fails to effectively increase its weight proportion; in the side-view angle collection scene, the visibility of the upper limb chain motion feature may decrease, but the existing method does not adaptively adjust the feature priority.

[0080] The application further proposes a process of dynamic weight distribution, including: when the light intensity is lower than a preset threshold, increasing the weight proportion of the skeletal point motion trajectory generated by the depth sensor in the local difference feature vector; when the collection viewing angle deviates from the normal viewing angle by more than a preset range, enhancing the priority of the upper limb chain motion feature in the global difference feature vector.

[0081] The preset threshold is determined based on the inflection point of the matching error curve of skeletal point trajectory data and contour image features under different light intensities. The weight proportion adjustment is realized by modifying the scaling factor of the feature vector, specifically multiplying the skeletal point trajectory feature component by a preset proportion coefficient. The viewing angle deviation range is defined as triggering priority adjustment when the normal plane angle exceeds θ1 degrees, and the sorting position of the upper limb chain feature in the global feature is improved by 2-3 positions through the feature vector sorting algorithm. The above adjustment operation is triggered by real-time monitoring of the environmental parameter encoding vector E, wherein the light intensity parameter is updated every 200 milliseconds, and the viewing angle parameter is updated every 500 milliseconds.

[0082] Specifically, in low light conditions, the ambient light intensity is continuously monitored by the light-sensitive sensor, and when the light values of three consecutive sampling periods are detected to be lower than the preset threshold, 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 remains the original coefficient. In the case of a view angle deviating from the scene, when the depth sensor detects that the angle between the target object centroid and the collection plane normal reaches θ2 degrees, the upper limb chain motion features in the global difference feature vector are reordered to the top three positions of the feature sequence, and their participation in the feature matching stage is improved through the spatial attention mechanism. For example, in the gait cycle collected at a side view angle, the amplitude of the upper limb swing is difficult to accurately capture due to perspective distortion, so the stable features of the lower limb chain motion trajectory are preferentially used, and the ordering priority of the upper limb chain features is enhanced to compensate for the information loss caused by the change in the view angle. The dynamic adjustment process is realized through a programmable logic controller, which can quickly update the weight coefficients in real time and ensure the rapid adaptability of the identity discrimination features when the environmental parameters change.

[0083] As a preferred embodiment, the scheme of the present application is implemented as follows:

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

[0085] First, set the light intensity threshold to 100 lux. When the ambient light intensity is lower than 100 lux, the weight proportion 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 reduced.

[0086] Second, set the collection view angle threshold to 30°, and when the collection view angle deviates from the front view angle by more than 30°, the priority of the upper limb chain motion features in the global difference feature vector is enhanced. Specifically, the weight coefficient of the upper limb chain motion features is increased from the default value, while the weight coefficient of the lower limb chain motion features is correspondingly reduced.

[0087] Through the above technical scheme, the present application can adaptively adjust the feature weight according to the actual environmental conditions, rely more on depth information in low light environments, and make more use of upper limb features in non-front view angles, thereby improving the robustness and accuracy of the gait recognition system in complex environments. Thus, the problem of unstable recognition performance of traditional fixed weight methods in different environments is effectively overcome, and environmental adaptability of gait recognition is achieved.

[0088] In some of the above schemes of the present application, when the illumination intensity or the collection angle in the environmental parameters changes significantly, the traditional single matching strategy is difficult to balance the recognition accuracy and the computational efficiency. Specifically, under weak light or side view angle conditions, global gait features are easily disturbed by noise, while local limb features may have data missing or distortion, resulting in reduced reliability of full feature matching.

[0089] The present application further proposes to select a matching strategy according to the environmental parameters, enabling three-level full feature matching under normal view angle and strong light conditions; enabling lower limb feature priority matching under side view angle or weak light conditions.

[0090] Among them, the three-level full feature matching realizes progressive feature comparison through a multi-stage screening mechanism, and the lower limb feature priority matching improves the utilization rate of key features through feature dimension screening. Both are dynamically associated with the environmental parameter encoding vector, and the strategy switching is triggered through a preset threshold interval. The selection logic of the matching strategy forms a data-level cooperation with the sampling density adjustment in the space-time segmentation processing, ensuring the best adaptation of the input features and the matching algorithm.

[0091] Specifically, when the environmental parameters meet the normal view angle and the illumination intensity reaches the strong light level, the system performs three-level full feature matching. The first level matching selects a candidate feature subset from the gait feature library through cosine similarity calculation, preliminarily narrowing down the comparison range; the second level matching uses dynamic time warping algorithm to align and compensate the gait cycle phase of the candidate features, eliminating the trajectory deviation error caused by the difference in step frequency; the third level matching generates a comprehensive matching degree index by weighted fusion of cosine similarity and trajectory coincidence degree, and finally selects the feature corresponding to the highest value as the determination result. When the environmental parameters enter the side view angle or weak light interval, the system switches to the lower limb feature priority matching mode. At this time, the gait cycle length and single step horizontal displacement are first extracted as global stability indicators for preliminary screening, and then cosine similarity calculation is performed on the lower limb chain related feature components, and the lower limb motion features less affected by environmental interference are preferentially used to complete accurate matching. Through the dynamic selection mechanism of the matching strategy, the system can adaptively adjust the feature comparison dimension and the allocation of computing resources under complex environments, effectively balancing the recognition accuracy and real-time requirements.

[0092] As a preferred embodiment, the scheme of the present application is implemented as follows: when the collection view angle is located in the front view angle interval and the light intensity is in the strong light level, three-level full feature matching is activated. In the first matching stage, the system calculates the cosine similarity of the identity distinguishing feature and all samples in the feature library, and selects the top P3% similar samples to form a candidate subset; in the second matching stage, the candidate subset is subjected to dynamic time warping processing, and the trajectory coincidence degree of the lower limb swing angle and the trunk tilt angle is calculated through sliding window comparison of the gait cycle phase; in the third matching stage, the cosine similarity and the trajectory coincidence degree are fused according to the preset weight ratio, and the sample with the highest comprehensive score is selected as the determination result. When the collection view angle belongs to the side view angle interval or the light intensity is in the weak light level, lower limb feature priority matching is enabled: first, the lower limb chain joint motion frequency component in the identity distinguishing feature and the gait cycle length parameter are extracted; the absolute difference between the gait cycle length and the feature library samples is calculated for preliminary screening, and the samples with a difference less than S1 seconds are retained; then, only the lower limb chain joint motion frequency component is subjected to cosine similarity calculation, and the sample with the highest similarity is selected as the final determination result.

[0093] Through the above technical scheme, the present application solves the problem of the existing method that the single matching strategy in complex environment leads to a decrease in recognition accuracy. By dynamically selecting a feature matching strategy, full use is made of complete feature information under strong light conditions to improve recognition accuracy, and the robustness of lower limb feature matching is focused on under weak light or side view angle conditions, effectively reducing the influence of environmental interference on the recognition result. The scheme significantly improves the gait recognition stability under different light conditions and collection view angles while maintaining the algorithm generalization ability.

[0094] In some schemes of the present application described above, the dynamic weight distribution process adjusts the feature weight according to the environmental parameters, but in the actual matching process, full feature matching under different environmental conditions may introduce noise interference, leading to a decrease in matching efficiency and an increase in misjudgment rate. For example, in low light or side view angle conditions, the upper limb feature extraction may contain more errors, and if full feature matching is still used, the recognition reliability will be reduced.

[0095] The present application further proposes to select a matching strategy according to the environmental parameters, to enable three-level full feature matching under front view angle and strong light conditions, and to enable lower limb feature priority matching under side view angle or weak light conditions.

[0096] In the three-stage full feature matching, the similarity calculation is divided into three stages. Firstly, the cosine similarity is used to establish a candidate subset. Secondly, the dynamic time warping is used to improve the accuracy of phase alignment. Finally, the multi-dimensional similarity is fused to obtain the result. The lower limb feature priority matching is implemented in two stages. In the first stage, the gait cycle parameters are used for coarse screening. In the second stage, the lower limb features are focused on for accurate comparison. The correspondence between the environmental parameters and the matching strategy is realized through preset condition judgment. When the detected light intensity is lower than the preset threshold or the viewing angle deviates from the normal viewing angle by more than the preset range, the matching strategy is automatically switched.

[0097] Specifically, in the normal viewing angle and sufficient light, the system performs the three-stage full feature matching process. In the first stage, the cosine similarity of the to-be-identified feature and all samples in the database is calculated, and the candidate set with a similarity higher than the set threshold is selected. In the second stage, the dynamic time warping is performed on the samples in the candidate set to eliminate the influence of gait cycle phase difference on the trajectory coincidence calculation. In the third stage, the results of the first two stages are fused according to the preset weight to generate a comprehensive matching degree ranking. In the side viewing angle or weak light condition, the lower limb chain motion features and gait cycle parameters are preferentially extracted. In the first stage, the gait cycle length and single-step displacement parameters are used for rapid screening. In the second stage, only the cosine similarity of the lower limb motion features of the candidate samples is calculated. This strategy effectively avoids the error interference of the upper limb features in adverse conditions, such as the positioning deviation of the upper limb joint points in the weak light. By preferentially matching the lower limb features, the recognition reliability can be improved.

[0098] As a preferred embodiment, the scheme of the present application is implemented as follows: in the normal viewing angle scene with light intensity of 1200 lux, the three-stage full feature matching is triggered to execute. In the first stage of matching, the cosine similarity of several samples in the pre-stored gait feature library and the identity discrimination feature is calculated, and the candidate samples with a similarity higher than the threshold L1 are selected to form a subset. In the second stage of matching, the dynamic time warping is performed on the gait cycle sequence of the candidate samples, and the foot contact point in the gait phase is used as the alignment reference to calculate the coincidence degree of the skeletal point trajectory on the time axis, wherein the trajectory coincidence degree threshold is set to L2. In the third stage of matching, the cosine similarity and the trajectory coincidence degree are weighted and summed according to the preset weight ratio, and finally the sample with the highest comprehensive matching degree is selected as the determination result. In the weak light condition, when the environmental light intensity decreases to 50 lux and the collection viewing angle is 45° to the side, the lower limb feature priority matching strategy is started. Firstly, the samples with a gait cycle length mean of S2 seconds and a single-step horizontal displacement in the range of [M1, M2] meters are extracted from the pre-stored feature library, and the candidate samples are selected. Then, the lower limb chain joint angle change sequence of the candidate samples is extracted, and the cosine similarity with the lower limb feature in the current identity discrimination feature is calculated, and finally the sample with the highest similarity is selected to complete the identity determination.

[0099] By the technical solution, the application effectively solves the problem of insufficient gait feature matching precision in a complex environment. In low light or side view conditions, by preferentially matching lower limb movement features, local feature distortion caused by light interference can be avoided, and the relative stability of human lower limb movement is used to improve recognition reliability. When the environmental conditions are good, the multi-level feature fusion matching mechanism verifies through time and space, significantly reducing the probability of false matching between similar gaits. This method dynamically adjusts the matching strategy according to different environments, ensuring the robustness of the recognition system and optimizing the allocation of computing resources.

[0100] Embodiment two:

[0101] As shown in the gait recognition system based on local and global difference modeling network, comprising: Figure 3 A data acquisition module is configured to acquire a gait contour sequence of a target object and an environmental parameter;

[0102] A data processing module is configured to adjust the segmentation granularity according to the environmental parameter, perform spatiotemporal segmentation processing on the gait contour sequence, and generate local limb movement trajectory segments and global gait cycle sequences;

[0103] A feature extraction module is configured to perform frequency domain feature transformation on the local limb movement trajectory segments to obtain local difference feature vectors, and perform spatial alignment coding on the global gait cycle sequences to obtain global difference feature vectors;

[0104] A feature fusion module is configured to perform cross-scale interaction calculation on the local difference feature vectors and the global difference feature vectors to generate fusion difference features with spatiotemporal consistency, and perform dynamic weight distribution on the fusion difference features according to the environmental parameter to obtain optimized identity discrimination features;

[0105] A strategy selection module is configured to select a matching strategy according to the environmental parameter; the matching strategy includes three-level full feature matching and lower limb feature priority matching;

[0106] A feature matching module is configured to perform similarity matching between the identity discrimination features and a pre-stored gait feature library according to the corresponding matching strategy, and output an identity determination result of the target object.

[0107] The above content is only an example and description of the structure of the application. Those skilled in the art can make various modifications or supplements to the described specific embodiments or use similar ways to replace them, as long as they do not deviate from the structure of the application or exceed the scope defined by the present claims, and should belong to the protection scope of the application.

[0108]

[0109] ​In the description of the specification, reference to "one embodiment", "an example", "a specific example" or the like means that a particular feature, structure, material or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of the application. The appearances of the phrases "in one embodiment", "an example", "a specific example" or the like in various places in the specification are not necessarily referring to the same embodiment or example. Furthermore, the particular features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0110] The preferred embodiments of the application disclosed above are only to help explain the application. The preferred embodiments do not describe all the details of the application and limit the application to the specific embodiments. Obviously, many modifications and variations can be made in light of the contents of the specification. The specification selects and specifically describes these embodiments in order to better explain the principles and practical application of the application, so that those skilled in the art can well understand and utilize the application. The application is limited only by the claims and their full scope and equivalents.

Claims

1. A gait recognition method based on a local and global difference modeling network, characterized in that: The following steps are involved: Obtaining the target object's gait profile sequence and environmental parameters; Adjusting the segmentation granularity according to the environmental parameters, performing spatiotemporal segmentation processing on the gait profile sequence to generate local limb motion trajectory segments and a global gait cycle sequence; Performing frequency domain feature transformation on the local limb motion trajectory segments to obtain local difference feature vectors, and performing spatial alignment encoding on the global gait cycle sequence to obtain a global difference feature vector; Performing cross-scale interactive calculation on the local difference feature vector and the global difference feature vector to generate a fused difference feature with spatiotemporal consistency; Dynamically weighting the fused difference features according to the environmental parameters to obtain optimized identity discrimination features; Selecting a matching strategy based on the environmental parameters; the matching strategy includes three-level full feature matching and lower limb feature priority matching; The identity discrimination feature is matched with the pre-stored gait feature library for similarity according to the corresponding matching strategy, and the identity determination result of the target object is output.

2. The gait recognition method based on local and global difference modeling network according to claim 1, characterized in that: The process of obtaining the gait profile sequence includes: Acquiring a gait profile sequence using a multi-source sensing device, wherein the multi-source sensing device includes at least a visible light camera and a depth sensor; The visible light camera is used to collect RGB gait image sequences, while the depth sensor is used to obtain the 3D skeletal point motion trajectory. When performing contour extraction on the RGB gait image sequence, an adaptive illumination compensation algorithm is used to eliminate pixel distortion in low-light areas; The kinematic constraint check is performed on the motion trajectory of the three-dimensional skeleton points to eliminate abnormal data points that exceed the threshold of human joint movement.

3. The gait recognition method based on local and global difference modeling network according to claim 2, characterized in that: The process of obtaining the environmental parameters includes: The environmental parameters include light intensity and acquisition viewing angle; Measure light intensity in real time through the built-in photosensor of the visible light camera; The angle between the center of mass of the target object and the normal of the acquisition plane is calculated by the depth sensor as the acquisition viewing angle; The light intensity is divided into three levels according to the preset light intensity threshold: strong light, normal light, and weak light; The acquisition viewing angle is divided into two azimuth intervals of front viewing angle and side viewing angle according to the preset angle threshold; A timestamp synchronization mechanism between environmental parameters and gait profile sequences is established to ensure the temporal consistency of parameter association.

4. The gait recognition method based on local and global difference modeling network according to claim 3, characterized in that: The process of the spatiotemporal segmentation processing includes: The gait profile sequence is divided into N phase intervals, each phase interval corresponds to the spatiotemporal segmentation parameters of the foot contact, leg swing, and take-off phases; In a single phase interval, the human body is divided into three local areas: upper limb chain, trunk chain, and lower limb chain through the kinematic chain model. When sampling the motion trajectory of each local area, a time window sliding mechanism is used to ensure that the overlap rate of adjacent trajectory segments is not lower than the preset threshold. The process of adjusting the segmentation granularity according to the environmental parameters includes: A fine-grained segmentation mode is used under strong light conditions to increase the sampling density of the upper limb chain motion trajectory; A coarse-grained segmentation mode is used under weak light conditions to retain the sampled data of the lower limb chain motion trajectory; The sampling data of the lower limb chain motion trajectory is retained when viewed from the side.

5. The gait recognition method based on local and global difference modeling network according to claim 4, characterized in that: The frequency domain feature transformation process includes: Perform short-time Fourier transform on the local limb motion trajectory segments to extract the motion frequency distribution characteristics of each joint point; Performing dimensionality reduction processing on the motion frequency distribution characteristics by using principal component analysis, and retaining feature components whose variance contribution rate exceeds a preset threshold; The global gait cycle sequence is divided into three-dimensional space grids, and the density gradient change characteristics of the skeleton points in each grid unit are calculated. The calculation method of the density gradient change characteristics includes: Decompose the three-dimensional space grid into horizontal slice units and vertical slice units along the motion direction; Calculate the spatial distribution discreteness of the bone points in each slice unit and generate the density gradient change curve; The density gradient change curves of adjacent slice units are differentially calculated, and the curvature mutation points are extracted as feature marks.

6. The gait recognition method based on local and global difference modeling network according to claim 5, characterized in that: Also includes: Establish a dynamic mapping relationship between local difference feature vectors and global difference feature vectors, and calculate the contribution weight of each local area to the global feature through the spatial attention mechanism; The cross-scale interactive computation also includes an environment-aware attention mechanism: Generate the gating coefficient according to the environmental parameter encoding vector E, and the calculation process satisfies: Aenv=σ(Wl*L+Wg*G+We*E), where L is the local difference feature vector, G is the global difference feature vector, Wl, Wg, 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: Perform logarithmic transformation on the light intensity to obtain the normalized light encoding value; Perform sine and cosine encoding on the captured viewing angle to generate an orientation feature vector; The two types of encoding vectors are concatenated and then reduced in dimension through a fully connected layer to form the environment parameter encoding vector E.

7. The gait recognition method based on local and global difference modeling network according to claim 6, characterized in that: The process of dynamic weight allocation includes: When the light intensity is lower than a preset threshold, the weight of the skeleton point motion trajectory generated by the depth sensor in the local difference feature vector is increased; When the acquisition viewing angle deviates from the normal viewing angle by more than a preset range, the priority of the upper limb chain motion feature in the global difference feature vector is enhanced.

8. The gait recognition method based on local and global difference modeling network according to claim 7, characterized in that: The process of selecting a matching strategy based on environmental parameters includes: Enable three-level full feature matching under normal viewing angle and strong lighting conditions; Enables lower limb feature-prioritized matching in side-view or low-light conditions.

9. The gait recognition method based on local and global difference modeling network according to claim 8, characterized in that: The three-level full feature matching includes: First-level matching: Calculate the cosine similarity between the identity discriminant features and all samples in the pre-stored gait feature library to filter out the candidate feature subset; Second-level matching: performing dynamic time warping on the candidate feature subset and calculating the trajectory overlap after gait cycle phase alignment; Third-level matching: The cosine similarity and trajectory overlap are combined to generate a comprehensive matching degree, and the feature with the highest comprehensive matching degree is selected as the judgment result; The process of preferentially matching lower limb features includes: Extract lower limb chain-related feature components from the local difference feature vector, and extract gait cycle length and single-step horizontal displacement from the global difference feature vector; First-level matching: Calculate the similarity between the gait cycle length and the single-step horizontal displacement and all samples in the pre-stored gait feature library to filter out the candidate feature subset; Second-level matching: performing cosine similarity calculation on the candidate feature subset and the limb chain-related feature components, and the feature with the highest cosine similarity is used as the judgment result.

10. A gait recognition system based on a local and global difference modeling network, characterized by: A gait recognition method based on a local and global difference modeling network as claimed in any one of claims 1 to 9 is used, comprising: A data acquisition module, used to obtain the gait profile sequence and environmental parameters of the target object; a data processing module, configured to adjust the segmentation granularity according to the environmental parameters, perform spatiotemporal segmentation processing on the gait profile sequence, and generate local limb motion trajectory segments and a global gait cycle sequence; a feature extraction module, configured to perform frequency domain feature transformation on the local limb motion trajectory segments to obtain local difference feature vectors, and perform spatial alignment encoding on the global gait cycle sequence to obtain global difference feature vectors; A 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 a fused difference feature with spatiotemporal consistency; and dynamically assign weights to the fused difference feature according to the environmental parameters to obtain an optimized identity discrimination feature; A strategy selection module, configured 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 feature and the pre-stored gait feature library according to the corresponding matching strategy, and output the identity determination result of the target object.

Images