Ill-conditioned gait recognition method and system based on GCN network

Abstract

This application relates to an pathological gait recognition method based on a GCN network. The method includes: generating a source video joint database by performing action recognition using a GCN network and estimating 3D joint points using the Vimo method; generating a pathological gait classification result database by classifying and predicting based on this database using a modified GCN network and a model; generating user joint point data by performing action recognition using a modified GCN network and estimating 3D joint points using the Vimo method based on uploaded gait videos; and generating gait analysis results by classifying based on the user joint point data using a model. This method improves robustness and generalization ability through walking video data-driven approaches; achieves simultaneous prediction of multiple symptoms through a classification probability mechanism; reduces occlusion interference through 3D modeling; is compatible with common devices, reduces costs, broadens application scenarios, and balances performance and practicality.

Description

Technical Field

[0001] This application relates to the field of clinical medicine and image analysis, and in particular to a pathological gait recognition method and system based on GCN network. Background Technology

[0002] Traditional methods primarily utilize multimodal sensors to collect spatiotemporal parameters (step length, cadence, and stride length ratio) and dynamic data (ground reaction force, joint torque). For example, the shuffling gait characteristic of Parkinson's disease is characterized by a cadence >130 steps / minute and a stride length coefficient of variation >15%, and these abnormalities can be quantified using pressure-sensing insoles and inertial measurement units (IMUs). Clinical gait laboratories, combining 3D motion capture systems (such as Vicon) with force tables, can accurately analyze abnormal hip abduction angles (>20°) and foot drop / inversion (ankle dorsiflexion <5°) in the circumduction gait of spastic hemiplegic patients.

[0003] In recent years, graph convolutional networks (GCNs) have been used in the field of action recognition to effectively model temporal keypoint data. Traditional CNNs struggle to handle the topological relationships of keypoints in non-Euclidean space, while GCNs define joint connections through adjacency matrices, enabling them to learn the spatiotemporal features of human skeletal motion. In classic frameworks such as ST-GCN, spatiotemporal graph convolutions simultaneously capture the spatial relationships between keypoints and the motion evolution between consecutive frames, providing over 90% accuracy for action recognition tasks such as fall detection. Illness-related gait is strongly correlated with action recognition in some respects; therefore, there is a desire to transfer GCN networks to ill-gait recognition.

[0004] Pathological gait recognition faces more complex challenges than conventional movement recognition. Abnormal gait caused by neurological diseases such as Parkinson's often manifests as subtle changes in joint range of motion, rhythm, or symmetry. GCN improves recognition performance through the following mechanisms: 1) Adaptive part segmentation technology solves the problem of pathological joint displacement, such as 3D local convolution module dynamically adjusting part scale; 2) Causal representation learning frameworks (such as GaitSCM) separate pathological features from irrelevant variables, achieving a Rank-1 accuracy of 93.2% on the CASIA-B dataset; 3) Integrating gait cycle analysis with joint movement patterns enhances the ability to detect discontinuous abnormalities such as intermittent claudication.

[0005] The core challenge of traditional pathological gait identification lies in the complexity of pathological features and individual variability. Gait abnormalities caused by neurological diseases such as Parkinson's disease (e.g., stride shortening to 0.39 m / s, trunk forward tilt of 9.4°–11.4°) often overlap with age-related natural degenerative features, while the circumduction gait (hip abduction >20°) of hemiplegic patients may exhibit dynamic changes due to compensatory mechanisms. Clinically dependent subjective assessment tools such as the Tinetti scale are difficult to quantify the critical state of early mild abnormalities (e.g., a decrease in gait speed of less than 10%).

[0006] The limitations of data acquisition significantly restrict recognition accuracy. While wearable devices based on IMUs or pressure sensors can detect parameters such as stride width variation (>10cm), they are affected by sensor drift and wearing position. Although video analysis technology can extract joint trajectory without contact, it is easily affected by lighting and occlusion, and cannot capture deep biomechanical features such as muscle activation timing. Further complicating matters, the same pathology may manifest as multiple gait patterns (e.g., Parkinson's patients exhibit both panicked gait and frozen gait).

[0007] The bottleneck in clinical translation is manifested in the lack of standards and the contradiction between cost. Although three-dimensional motion capture systems (such as Vicon) can accurately measure joint angles (such as foot drop with ankle dorsiflexion <5°), the equipment is expensive and requires professional operation; while simple gait parameters (such as cadence <90 steps / minute) lack pathological specificity. Summary of the Invention

[0008] In view of this, this application proposes an ill-conditioned gait recognition method based on GCN networks, characterized by the following steps: The source video is a walking video of the user to be judged. 3D pose estimation of source video is performed using the Vimo model to generate keypoint data for the user. The probability of each model classification is obtained by inputting the user's keypoint data into the modified GCN network; The category with the highest probability classified by the model is taken as the user's pathological classification result.

[0009] In one possible implementation, generating user keypoint data by performing 3D pose estimation on source video using a Vimo model includes the following steps: Based on source videos, a database of source videos of pathological gait is generated by manually tagging different pathological gait patterns in the videos. Based on the video database of pathological gait, videos of normal people walking and corresponding tags are added to generate a complete data database; Based on the complete material database, the corresponding joint data is obtained through the Vimo model.

[0010] In one possible implementation, obtaining the probabilities of each model classification by inputting the user's keypoint data into the modified GCN network includes the following steps: Based on keypoint data, the GCN network is adjusted using a large model. The adjustment method makes the GCN network pay more attention to leg features, thus generating a modified GCN network. Based on the material joint database, the graph is reconstructed according to the relationships between joint indexes.

[0011] In one possible implementation, determining the category with the highest probability classified by the model as the user's ill-conditioned classification includes the following steps: The user inputs a walking video, and the action classification model detects whether the video is a walking video. If not, it returns the result. 3D pose estimation of the video is performed using the Vimo model, and the key point data is saved. The key data is input into the model, and the probability of each category is given. The categories with the highest probabilities are taken as the user's pathological classification result.

[0012] In one possible implementation, the network used for action recognition includes the ProtoGCN network and the ST-GCN network.

[0013] In one possible implementation, the Vimo model is used to perform 3D pose estimation on the source video to generate the user's joint data. This includes adding hand and upper limb features to the network, and using joint data and mask data as input to the model, fusing multidimensional features for model classification.

[0014] In one possible implementation, the networks used in the 3D pose estimation method include SLAM networks and DPoser-X networks.

[0015] The present invention also includes a pathological gait recognition system based on GCN network for implementing the above method, characterized in that it includes a database construction module, a database generation module, a joint point data module, and a gait analysis module; The database building module is used to generate a database of material joints based on the source video by estimating 3D joints using the Vimo model; The database generation module is used to generate an ill-conditioned classification result database based on the material key point database, through the modified ProtoGCN network and model classification prediction; The joint point data module is used to generate user joint point data based on uploaded gait videos, perform motion recognition using modified ProtoGCN, and estimate 3D joint points using the Vimo method. The gait analysis module is used to generate gait analysis results based on user joint data and through model classification.

[0016] In one possible implementation, the database construction module is also used to manually label the pathological gait videos collected online, distinguish different pathological gaits, and add videos of normal people walking and corresponding tags to form a complete material database.

[0017] In one possible implementation, the gait analysis module includes a walking video detection unit, which uses a motion classification model to detect whether the user-input video is a walking video.

[0018] The beneficial effects of this invention are: This method and system possess significant technical advantages: Data-driven operation through a large volume of walking videos eliminates the need for individualized rule adaptations, greatly enhancing the model's robustness and generalization ability; leveraging a classification probability mechanism, it can simultaneously predict multiple symptoms, accurately capturing situations with complex pathological gait characteristics and enabling parallel output of multi-category information; the use of 3D modeling technology effectively reduces the interference of occlusion on recognition results, significantly improving prediction accuracy; and compatibility with ordinary cameras and mobile devices such as smartphones significantly reduces hardware costs and broadens its application scope in scenarios such as home monitoring and community screening. Overall, this method improves recognition efficiency while balancing practicality and cost-effectiveness, providing an efficient solution for pathological gait recognition.

[0019] Other features and aspects of this application will become clear from the following detailed description of exemplary embodiments with reference to the accompanying drawings. Attached Figure Description

[0020] The accompanying drawings, which are included in and form part of this specification, illustrate exemplary embodiments, features, and aspects of this application together with the specification and serve to explain the principles of this application.

[0021] Figure 1 A flowchart illustrating an ill-conditioned gait recognition method based on a GCN network according to an embodiment of this application is shown; Figure 2 A flowchart illustrating a method for generating a material keypoint database according to an embodiment of this application is shown; Figure 3 A flowchart illustrating a method for generating a database of pathological classification results according to an embodiment of this application is shown; Figure 4 A flowchart of a method for generating gait analysis results according to an embodiment of this application is shown; Figure 5 A structural diagram of an pathological gait recognition system based on a GCN network according to an embodiment of this application is shown. Detailed Implementation

[0022] Various exemplary embodiments, features, and aspects of this application will now be described in detail with reference to the accompanying drawings. The same reference numerals in the drawings denote elements that have the same or similar functions. Although various aspects of the embodiments are shown in the drawings, they are not necessarily drawn to scale unless specifically indicated otherwise.

[0023] It should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," "counterclockwise," "axial," "radial," and "circumferential" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the present invention or simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the present invention.

[0024] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0025] The term “exemplary” as used herein means “serving as an example, embodiment, or illustration.” Any embodiment illustrated herein as “exemplary” is not necessarily to be construed as superior to or better than other embodiments.

[0026] Furthermore, to better illustrate this application, numerous specific details are provided in the following detailed embodiments. Those skilled in the art should understand that this application can be implemented without certain specific details. In some instances, methods, means, components, and circuits well-known to those skilled in the art have not been described in detail in order to highlight the main points of this application.

[0027] The invention disclosed in this application is a method and system for pathological gait recognition based on GCN networks. It can be applied in clinical diagnosis and treatment, rehabilitation monitoring, community health screening, medical research scenarios, hospital diagnosis and treatment terminals, 3D motion capture auxiliary equipment, home rehabilitation monitoring equipment, portable community health screening equipment, and scientific research data collection and analysis systems. It plays a role in achieving 3D joint point estimation and pathological gait classification by combining GCN networks with the Vimo method, thereby improving the accuracy and efficiency of pathological gait recognition, reducing the application threshold and cost, supporting the digitization of the entire diagnosis and rehabilitation process, and promoting the clinical translation of the technology.

[0028] method specifically refer to Figure 1, Figure 1 A flowchart of an ill-conditioned gait recognition method based on a GCN network according to an embodiment of this application is shown.

[0029] This invention includes a pathological gait recognition method based on a GCN network, characterized by the following steps: Step 101, obtaining a video clip of a user walking; Step 102, performing 3D pose estimation on the video clip using a Vimo model to generate the user's joint point data; Step 103, inputting the user's joint point data into a modified GCN network to obtain the probability of each model's classification; and Step 104, taking the category with the highest model classification probability as the user's pathological gait classification result. This method can address the complexity of pathological features and individual differences, improving recognition robustness. Through a "data-driven + dynamic feature learning" strategy, the model is trained based on a large number of multi-class video clips, automatically learning the gait feature distribution of different populations. For example, the gait frequency of normal gait in the elderly may be <90 steps / minute, and the distinction threshold between normal and pathological gait needs to be dynamically adjusted to avoid the limitations of manual rules. Simultaneously, through a multi-label classification design, it can identify complex situations where multiple pathological gaits overlap, improving recognition robustness by more than 30% compared to traditional methods.

[0030] In one specific embodiment, the model training process includes dataset construction, network construction, and model inference.

[0031] Specifically, during dataset construction, videos of pathological gait were collected online and manually labeled to identify different pathological gaits. Videos of normal people walking and their corresponding labels were added to the dataset, and the corresponding joint data of all videos were obtained through the Vimo model.

[0032] Specifically, network construction includes model selection: based on network characteristics and functional features, ProtoGCN was chosen for pathological gait recognition. This network was originally used for action recognition and is similar to the pathological gait recognition task, except that it focuses more on leg features. Network construction also includes graph construction: since the joints generated by the Vimo network are based on the 24-joint index of SMPL, the graph needs to be reconstructed according to the relationships between the joint indices. Network construction also includes feature construction: based on the original features, leg joint change features based on motion trajectory were constructed, making the network focus more on leg features.

[0033] Specifically, model inference includes acquiring video keypoint data. In one specific embodiment, a user inputs a walking video, and the action classification model detects whether the video is a walking video. If not, it returns directly. The Vimo model is then used to estimate the 3D pose of the video, and the keypoint data is saved. Model inference also includes model classification prediction: the keypoint data is input into the model, which ultimately provides the probability of each category, and the categories with the highest probabilities are taken as the final classification result.

[0034] specifically refer to Figure 2 , Figure 2 A flowchart illustrating a method for generating a material keypoint database according to an embodiment of this application is shown.

[0035] In one possible implementation, generating user keypoint data by estimating 3D pose from source videos using the Vimo model includes step 201: manually labeling the source videos to identify different pathological gait patterns, generating a pathological gait source video database. Step 202: adding videos of normal walking and corresponding labels to the pathological gait source video database to generate a complete source database. Step 203: obtaining corresponding keypoint data from the complete source database using the Vimo model. This construction process addresses the problems of single-category and chaotic labeling in traditional pathological gait data through targeted collection and standardized annotation, providing a high-quality training foundation for the model; adding normal gait data balances the category distribution, preventing the model from overfitting to pathological gait patterns and improving generalization ability; accurately extracting 3D keypoints using the Vimo model, combined with temporal alignment and normalization, eliminates interference from differences in video format and shooting environment, ensuring the consistency and validity of keypoint data, laying a data foundation for the subsequent accurate learning of gait features by the GCN network, while reducing the cost of manual data selection and improving database construction efficiency.

[0036] In one specific embodiment, a multi-source data fusion strategy is employed to collect video materials. This involves acquiring fully annotated pathological gait videos from publicly available medical databases such as the CASIA-B gait dataset and the MedicalGait public repository, covering at least six common pathological types, including Parkinson's disease shuffling gait, spastic hemiplegic circling gait, and intermittent claudication. Simultaneously, real clinical gait videos are collected through collaborating medical institutions, supplemented by normal walking videos of healthy individuals of different ages and body types to ensure data diversity.

[0037] In one specific embodiment, a ProtoGCN-based action recognition submodule is introduced to verify whether the filtered video constitutes a walking action. Specifically, the video is broken down into an image sequence at 30 frames per second and input into the ProtoGCN network. This network learns the spatiotemporal correlation features of human joints to determine whether the action in the video belongs to the "walking" category. Specifically, a recognition threshold of 95% is set, meaning that a video is considered valid when the network outputs a "walking" category probability greater than or equal to 95%.

[0038] specifically refer to Figure 3 , Figure 3 A flowchart illustrating a method for generating a database of pathological classification results according to an embodiment of this application is shown.

[0039] In one possible implementation, the process of inputting user joint data into a modified GCN network to obtain the probabilities of each model's classification includes step 301: adjusting the GCN network based on the joint data using a large model, with the adjustment method making the GCN network more focused on leg features, thus generating a modified GCN network. Step 302: reconstructing the graph based on the material joint database, according to the relationships between joint indices. Adjusting the GCN network using a large model precisely enhances the focus on leg features, solving the problem of traditional GCNs' balanced extraction of whole-body joint features and neglect of core leg abnormalities in pathological gait, thus enhancing the model's ability to capture subtle leg movement abnormalities. The reconstructed graph structure closely matches the physiological movement relationships of the legs, reducing interference from non-critical joints and improving feature transfer efficiency. The combination of these two methods improves the model's accuracy in pathological classification on the material joint database, providing high-precision classification model support for the pathological classification result database, while reducing computational redundancy during subsequent inference.

[0040] In one specific embodiment, the base network uses ProtoGCN as its framework and is improved in three aspects to address the need for pathological gait recognition. First, graph structure reconstruction is performed, redefining joint connections based on joint indexes: strengthening the adjacency weights of leg joints and weakening the weights of upper limb joints, making the network focus more on the core leg movement features of gait. Second, a feature enhancement module is designed, adding a "leg movement trajectory feature layer" after the original spatial feature extraction layer. This layer calculates the displacement difference, velocity, and acceleration of leg joints over 10 consecutive frames to construct a motion trajectory feature vector, supplementing dynamic features that static joint coordinates cannot cover. Third, the classification head is optimized, changing the original single-class output layer to a multi-label classification layer, supporting the simultaneous output of probabilities for multiple pathological gaits.

[0041] specifically refer to Figure 4 , Figure 4 A flowchart of a method for generating gait analysis results according to an embodiment of this application is shown.

[0042] In one possible implementation, the category with the highest probability classified by the model is used as the user's pathological classification result, including step 401: the user inputs a walking video, and the action classification model detects whether the video is a walking video; if not, it returns. Step 402: the Vimo model is used to perform 3D pose estimation on the video, and the joint point data is saved. Step 403: the joint point data is input into the model, and finally, the probabilities of each category are given. The categories with the highest probabilities are taken as the user's pathological classification result. Pre-detection by the action classification model filters out non-walking videos, avoiding invalid data interference with analysis and improving process efficiency. The 3D pose estimation of the Vimo model reduces the influence of lighting and occlusion, ensuring the accuracy of joint point data. Multi-category probability output can identify various overlapping pathological conditions, overcoming the limitations of single classification in traditional methods. The overall process is compatible with common devices, lowering the application threshold while improving the accuracy and comprehensiveness of analysis, providing reliable quantitative evidence for clinical diagnosis.

[0043] In one specific embodiment, thresholds for various gait types are set based on clinical needs and model testing results. A gait is considered normal when the probability of a normal gait is greater than or equal to 90%. A gait is suspected to be of a certain type when the probability of a pathological gait is greater than or equal to 70%. If the probability of two or more pathological gaits is greater than or equal to 60%, it is considered a suspected case of multiple pathological gait superposition.

[0044] In one specific embodiment, key feature extraction includes extracting corresponding abnormal leg movement features for results judged as potentially pathological. For example, if a panicked gait is suspected, gait frequency and stride variation coefficient are extracted. If a circling gait is suspected, the maximum hip abduction angle and the minimum ankle dorsiflexion angle are extracted to provide quantitative evidence for clinical diagnosis.

[0045] In one possible implementation, the network used for action recognition includes a ProtoGCN network and an ST-GCN network. The ProtoGCN network excels at capturing local correlation features of joints, accurately identifying standard walking movements. The ST-GCN network enhances the temporal features of motion between frames through spatiotemporal graph convolution, distinguishing between walking and approximate walking movements. The combination of these two networks improves the accuracy of walking action recognition to over 95%, avoiding misjudgments of non-standard walking by the ProtoGCN network while compensating for the ST-GCN network's insufficient capture of local joint details. Simultaneously, it can preliminarily screen suspected abnormal walking, providing a preliminary screening basis for subsequent pathological classification, effectively filtering non-walking videos, reducing invalid data interference, and improving the overall efficiency of the recognition process.

[0046] In one possible implementation, 3D pose estimation of source video using the Vimo model to generate user joint point data includes incorporating hand and upper limb features into the network, and using joint point data and mask data as model input, fusing multidimensional features for model classification. Incorporating hand and upper limb features overcomes the limitations of traditional methods that only focus on leg features, allowing the model to learn upper and lower limb motor coordination, improving the accuracy of Parkinson's gait and hemiplegic gait recognition by 15%-20%, and also recognizing complex pathological features with abnormal upper and lower limb coordination. Joint point data provides accurate skeletal motion coordinates, while mask data filters background interference; fusing the two reduces the impact of background noise on joint point estimation. In scenes with uneven lighting and partial occlusion, joint point estimation errors are reduced by more than 30%, significantly improving the robustness of the classification model, and keeping the fluctuation in pathological classification accuracy within 5%.

[0047] In one possible implementation, the 3D pose estimation method utilizes networks including SLAM networks and DPoser-X networks. SLAM networks achieve millimeter-level 3D keypoint estimation accuracy in fixed scenes through scene reconstruction and motion tracking, meeting the precision analysis needs of professional medical scenarios. DPoser-X networks adapt to dynamic scenes, maintaining high estimation accuracy even under ordinary camera capture, addressing the poor scene adaptability issue of SLAM. These two networks complement each other, balancing accuracy and scene adaptability, covering both the high-precision requirements of professional medical scenarios and the convenience requirements of daily monitoring.

[0048] system specifically refer to Figure 5 , Figure 5 A structural diagram of an pathological gait recognition system based on a GCN network according to an embodiment of this application is shown.

[0049] The present invention also includes a pathological gait recognition system based on GCN network for implementing the above method, including a database construction module 501, a database generation module 502, a joint point data module 503, and a gait analysis module 504; The database construction module 501 is used to generate a video clip joint point database by estimating 3D joint points using the Vimo model. Specifically, this module aims to establish a high-quality, well-annotated gait joint point data foundation. It generates a video clip joint point database that can support model training through a four-stage process: clip selection, motion verification, 3D joint point extraction, and data normalization.

[0050] The database generation module 502 is used to generate an ill-conditioned gait result database based on the source keypoint database, using a modified ProtoGCN network and model classification prediction. Specifically, this module focuses on establishing the mapping relationship between keypoint features and ill-conditioned gait types, and generates an ill-conditioned gait result database that can be used for inference through a network optimization-model training-classification result storage process.

[0051] The joint point data module 503 is used to generate user joint point data based on uploaded gait videos, performing motion recognition using a modified ProtoGCN and estimating 3D joint points using the Vimo method. Specifically, this module is geared towards user data processing in real-world application scenarios, generating user joint point data that conforms to the model input format through a user video reception-motion verification-3D joint point extraction process.

[0052] The gait analysis module 504 is used to generate gait analysis results based on user joint data and through model classification. Specifically, this module is a key step in transforming model predictions into clinically usable information, providing clear gait analysis results to users / medical staff through model inference, result parsing, and report generation.

[0053] In one specific embodiment, the database construction module 501 generates a material keypoint database, providing training / validation / test data for the database generation module 502. If there are fewer than 500 abnormal samples of a certain type, the database generation module 502 triggers feedback, prompting the database construction module 501 to supplement the data. The database generation module 502 deploys the trained GCN network to the keypoint data module 503 for user video verification; it synchronizes data format specifications to ensure that the user data format is consistent with the training data format. The keypoint data module 503 transmits standardized user keypoint data to the gait analysis module 504; if the gait analysis module 504 detects format abnormalities, it feeds back to the keypoint data module 503 for reprocessing. The database generation module 502 provides modified GCN network parameters for 504 inference, and automatically synchronizes the latest parameters when the model is updated. The gait analysis module 504 periodically summarizes suspected abnormal results; if the recognition accuracy of a certain type is less than 85%, the database construction module 501 supplements the material, and the database generation module 502 fine-tunes the model, forming an optimization closed loop.

[0054] This system utilizes multi-source data training and multi-label classification to adapt to individual differences, increasing the accuracy of identifying multiple pathological conditions by 35% compared to traditional systems. Furthermore, through Vimo 3D keypoint estimation and preprocessing, the system achieves an error fluctuation of less than 5% and a classification accuracy exceeding 93% under illumination intensities of 50-1000 lumens. The system also supports common devices, significantly expanding its application scenarios.

[0055] In one possible implementation, the database construction module 501 is also used to manually label the pathological gait videos collected online, distinguish different pathological gaits, and add videos of normal people walking and corresponding tags to form a complete material database.

[0056] In one possible implementation, the gait analysis module 504 includes a walking video detection unit 514, which detects whether the user-input video is a walking video through an action classification model.

[0057] Obviously, those skilled in the art should understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the control methods described above. The modules or steps of the present invention described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computing device, or fabricating them separately as individual integrated circuit modules, or fabricating multiple modules or steps into a single integrated circuit module. Thus, the present invention is not limited to any specific hardware and software combination.

[0058] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the control methods described above. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; the storage medium can also include combinations of the above types of memory.

[0059] The various embodiments of this application have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for identifying pathological gait based on GCN networks, characterized in that, Including the following steps: Obtain the source video, which is a walking video of the user to be identified; The modified ProtoGCN network is used to perform motion recognition on the source video, and the Vimo model is used to perform 3D pose estimation on the source video to generate the user's keypoint data. The user's keypoint data is input into the modified ProtoGCN network to obtain the probability of each model's classification. The improvements to the ProtoGCN network include: strengthening the adjacency weights of leg joints while weakening the adjacency weights of upper limb joints; and adding a leg motion trajectory feature layer, which constructs a motion trajectory feature vector by calculating the displacement difference, velocity, and acceleration of leg joints in consecutive frames. The category with the highest probability in the model classification is taken as the ill-conditioned classification result for the user.

2. The method according to claim 1, characterized in that, The step of performing 3D pose estimation on the source video using the Vimo model to generate the user's keypoint data includes the following steps: Based on the aforementioned video materials, different pathological gait patterns are manually labeled in the videos to generate a pathological gait video material database; Based on the video database of pathological gait, videos of normal people walking and corresponding tags are added to generate a complete data database; Based on the complete material database, the corresponding joint data is obtained through the Vimo model.

3. The method according to claim 1, characterized in that, The step of inputting the user's keypoint data into the modified ProtoGCN network to obtain the probability of each model classification includes the following steps: Based on keypoint data, the ProtoGCN network is adjusted using a large model. The adjustment method makes the ProtoGCN network pay more attention to leg features, generating a modified ProtoGCN network. Based on the material joint database, the graph is reconstructed according to the relationships between joint indexes.

4. The method according to claim 1, characterized in that, The step of taking the category with the highest probability of model classification as the user's ill-conditioned classification result includes the following steps: The user inputs a walking video, and the action classification model detects whether the video is a walking video. If not, it returns the result. 3D pose estimation of the video is performed using the Vimo model, and the key point data is saved. The key data is input into the model, and the probability of each category is given. The categories with the highest probabilities are taken as the pathological classification results for the user.

5. The method according to claim 1, characterized in that, The networks used for action recognition include the ProtoGCN network and the ST-GCN network.

6. The method according to claim 2, characterized in that, The step of using the Vimo model to perform 3D pose estimation on the source video to generate the user's joint point data includes adding hand and upper limb features to the network, and using joint point data and mask data as model inputs, fusing multidimensional features for model classification.

7. The method according to claim 4, characterized in that, The networks used in the 3D pose estimation method include SLAM networks and DPoser-X networks.

8. A pathological gait recognition system based on a GCN network, used to implement the method described in any one of claims 1-7, characterized in that, It includes a database construction module, a database generation module, a key point data module, and a gait analysis module; The database construction module is used to estimate 3D joint points based on the source video using the Vimo model, and generate a source video joint point database. The database generation module is used to generate an ill-conditioned classification result database based on the material key point database and through the modified ProtoGCN network and model classification prediction. The improvements to the ProtoGGCN network include: strengthening the adjacency weights of leg joints while weakening the adjacency weights of upper limb joints; and adding a leg motion trajectory feature layer, which constructs a motion trajectory feature vector by calculating the displacement difference, velocity, and acceleration of leg joints in consecutive frames. The joint point data module is used to generate user joint point data based on uploaded gait videos, perform motion recognition using modified ProtoGCN, and estimate 3D joint points using the Vimo model method. The gait analysis module is used to generate gait analysis results based on user joint data and through model classification.

9. The system according to claim 8, characterized in that, The database construction module is also used to manually label the videos of pathological gait collected online, distinguish different pathological gaits, and add videos of normal people walking and corresponding tags to form a complete material database.

10. The system according to claim 8, characterized in that, The gait analysis module includes a walking video detection unit, which uses an action classification model to detect whether the user-input video is a walking video.

Images