Infant rehabilitation training auxiliary method and system capable of intelligently recognizing motion track

Abstract

The invention relates to the field of infant early-stage development and rehabilitation medicine, and particularly discloses an infant rehabilitation training auxiliary method and system capable of intelligently recognizing action tracks. According to the method, computer vision and artificial intelligence technologies are fused, a special infant human body key point detection model and a multi-target tracker are constructed, accurate capture and quantitative analysis of an infant movement track are achieved, intelligent comparison and evaluation are conducted in combination with a standard development milestone database, and finally a personalized training guidance scheme is generated. According to the method, a complete technical closed loop of'evaluation-quantification-comparison-guidance 'is established, traditional development evaluation depending on subjective experience is converted into an objective, accurate and quantifiable intelligent analysis process, evaluation efficiency and consistency are remarkably improved, real-time and personalized technical support can be provided for scientific intervention in a family scene, and the method is suitable for popularization and application. The development of the athletic ability of the infants is effectively promoted.

Description

Technical Field

[0001] This invention belongs to the field of early childhood development and rehabilitation medicine, specifically a method and system for assisting infant rehabilitation training by intelligently recognizing movement trajectories. Background Technology

[0002] Early gross motor development in infants and toddlers is a core component of their neurodevelopment. Milestones such as head lifting, rolling over, crawling, and walking are not only indicators of physical growth but also closely related to the formation of cognitive, perceptual, and social abilities. Clinically, developmental scales (such as the Peabody Developmental Motor Scale and the Alberta Infant Motor Scale) are commonly used for assessment. These scales are based on standardized observation, comparing infant movements to age-appropriate norms. With the development of computer vision technology, video-based behavioral analysis has been gradually applied to the field of sports science. Algorithms extract key points on the human body and analyze movement trajectories, providing the technological possibility for objectively quantifying movement patterns.

[0003] Currently, there are studies attempting to apply general human posture estimation models to motion capture for infants and toddlers. However, infants and toddlers differ significantly from adults in body proportions, movement patterns, and behavioral scenarios: their limbs are shorter, joint features are less pronounced, and their movements are rapid, unstable, and often accompanied by activities on non-supporting planes (such as crawling). Existing general models suffer from low confidence in keypoint detection, easy loss of keypoints, and significant jitter in infant and toddler scenarios, leading to trajectory breaks or drift. This directly results in large errors in subsequent motion feature extraction, making it impossible to support a stable assessment of the quality of fine motor skills. More importantly, existing technical solutions mostly remain at the level of "detection and trajectory mapping," lacking a complete framework for systematically and quantitatively comparing and analyzing time-series trajectory data with standard developmental milestones. This makes it difficult to automatically identify developmental characteristics and determine the degree of deviation from massive amounts of motion data, and even more difficult to generate personalized intervention recommendations with clinical guidance. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides an intelligent method and system for assisting infant and toddler rehabilitation training by recognizing movement trajectories. This addresses the issues in existing technologies where infant and toddler gross motor development assessments rely on subjective experience, have low automation levels, and lack precise assessments and closed-loop intervention guidance based on quantitative data.

[0005] A method for assisting infant and toddler rehabilitation training by intelligently recognizing movement trajectories includes the following steps:

[0006] S1. Data Acquisition and Model Preparation: Construct an image dataset of keypoint-annotated images of infants and toddlers, and use this dataset to train an end-to-end human keypoint detection model for infants and toddlers; at the same time, configure a multi-target keypoint tracker based on temporal prediction and data association.

[0007] S2. Real-time detection and trajectory tracking: Real-time video stream of the infant's activity area is acquired through a visual acquisition device; for each frame of the image, the trained keypoint detection model is used for detection, and the output includes the coordinates and confidence scores of keypoints including at least the head, shoulder, elbow, hip, knee, and ankle joints; then, the keypoint detection results are input into the multi-target keypoint tracker, and a unique trajectory ID is assigned to and maintained for each detected keypoint through cross-frame data association, generating continuous and stable time-series keypoint coordinate data;

[0008] S3. Motion trajectory generation and feature quantization: Based on the coordinate data of key points with the same trajectory ID, construct the motion trajectory of a specific body part or body center of gravity in two-dimensional or three-dimensional space; perform parameterized calculation on the motion trajectory, and extract one or more quantified features including trajectory path length, average motion speed, peak acceleration, motion smoothness, joint angle change range and motion cycle to form a quantized feature vector representing the current motion;

[0009] S4. Motion Analysis and Developmental Assessment: The quantitative feature vector is input into a preset motion analysis model and compared with the standard motion template features of the corresponding age in the standard infant gross motor development milestone database; the motion analysis model comprehensively evaluates the completion, standardization and deviation from the developmental standard of the current motion by calculating feature similarity or distance, and generates a motion assessment report containing quantitative scores and qualitative descriptions;

[0010] S5. Personalized guidance generation and interactive feedback: Based on the action assessment report, a targeted follow-up training guidance plan is generated by adaptive matching or combination from the preset training action library; through the graphical user interface, the assessment results and training guidance are fed back to the caregiver in real time in one or more ways, such as overlaying comparison animation, highlighting deviation areas, voice prompts and development progress charts.

[0011] Preferably, in step S2, the workflow of the multi-target keypoint tracker includes:

[0012] S21. Confidence Filtering: Set a confidence threshold and only send key point detection results with a confidence level higher than the threshold into the tracking process;

[0013] S22. Trajectory Initialization and Management: Assign a new trajectory ID to the first valid key point that appears in each frame, and build a trajectory dictionary to store its historical coordinate sequence using the trajectory ID as the key;

[0014] S23. State prediction and data association: For an existing trajectory, use a Kalman filter to predict the expected position in the current frame based on its historical coordinates; calculate the correlation between the predicted position and the actual detected key point position in the current frame, use the Hungarian algorithm for optimal matching, and update the coordinates of the corresponding trajectory if a successful match is achieved.

[0015] S24. Trajectory Lifecycle Management: Retain the count of existing trajectories that fail to match. If the count is not updated after a preset number of frames, the trajectory is determined to be terminated and deleted. For new detection points that do not match, initialize them as new trajectories.

[0016] Preferably, in step S1, constructing the infant key point labeled image dataset includes: collecting images of infant activities at different ages, in different scenes, and under different lighting conditions; using a labeling tool to accurately label the key points of the infant's body in the images; and dividing the dataset into a training set, a validation set, and a test set according to a certain ratio for model training and evaluation.

[0017] Preferably, in step S3, if the visual acquisition device is a depth camera or an RGB-D camera, the depth information is used to calculate the three-dimensional spatial coordinates of the key points, construct the three-dimensional motion trajectory, and perform feature quantization in the three-dimensional space.

[0018] Preferably, in step S4, the standard infant gross motor development milestone database is dynamically updated, and its update data comes from the statistical analysis results of trajectory feature vectors after anonymization of a large number of infant users.

[0019] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described herein.

[0020] An intelligent motion trajectory recognition auxiliary system for infant rehabilitation training, used to implement the above method, includes:

[0021] The data acquisition module, deployed in the infant and toddler activity area, includes at least one visual acquisition device for continuously acquiring video stream data;

[0022] An edge computing unit, connected to the data acquisition module, is used to run the lightweight key point detection model and perform preliminary detection and preprocessing.

[0023] A cloud-based data processing and analysis platform, communicatively connected to the edge computing unit, includes:

[0024] Trajectory tracking server: used to run the multi-target keypoint tracker, receive keypoint detection results and output time-series trajectory data;

[0025] Feature extraction and evaluation engine: used to perform motion trajectory generation, feature quantization, and comparative analysis with standard databases to generate motion evaluation reports;

[0026] Personalized recommendation engine: It has the built-in training action library and generates personalized training guidance plans based on the evaluation report;

[0027] The user interaction terminal is connected to the cloud data processing and analysis platform to receive and visualize the action evaluation report and training guidance plan, and to provide a human-computer interaction interface.

[0028] Preferably, the visual acquisition device integrates an infrared transmitter and a receiver to form a depth sensing module for acquiring depth image information of infants and young children.

[0029] Preferably, the system also includes a parent-side mobile application as an implementation of the user interaction terminal, which provides functions such as video viewing, evaluation report history, training plan push, and community communication.

[0030] Preferably, the cloud-based data processing and analysis platform adopts a microservice architecture, with the trajectory tracking server, feature extraction and evaluation engine, and personalized recommendation engine deployed as independent services and exchanging data through message queues.

[0031] Compared with the prior art, the present invention has the following beneficial effects:

[0032] By leveraging an end-to-end keypoint detection model and multi-target tracking algorithm, the movement trajectory of key points on an infant's body can be captured and continuously locked in milliseconds, and converted into accurately calculable kinematic parameters. This transforms previously vague qualitative observations into objective and reproducible quantitative data, laying a solid scientific data foundation for developmental assessment.

[0033] By constructing a standard movement database deeply linked to infant age and an intelligent comparative analysis model, the system achieves precise benchmarking of developmental milestones and personalized gap diagnosis. The system does not merely provide raw movement data; instead, it intelligently matches and calculates similarity between real-time extracted trajectory feature vectors and a massive database of standardized developmental templates. This allows the system to accurately determine the micro-level gap between an infant's current movements and the expected standards for their corresponding age, specifically identifying whether it's insufficient speed, limited joint range of motion, or poor coordination. This generates a clearly targeted assessment report, overcoming the limitations of traditional methods that can only provide general conclusions.

[0034] By generating real-time feedback and personalized training guidance based on assessment results, a complete intervention loop of "assessment-guidance-training-reassessment" is formed. Based on accurate gap diagnosis, the system can automatically match or dynamically combine the most targeted rehabilitation games or training programs from a training action library, and guide caregivers in real-time interaction through intuitive methods such as augmented reality overlay comparison, highlighting of deviation areas, and voice guidance. This not only enhances the fun and compliance of home training but also enables rehabilitation guidance to be conducted frequently and scientifically in daily scenarios, effectively promoting the development of infants' and toddlers' motor skills and providing a reliable technological tool for the early detection of developmental deviations. Attached Figure Description

[0035] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0036] Figure 2 This is a system architecture diagram of the present invention. Detailed Implementation

[0037] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0038] Example 1:

[0039] This embodiment provides an intelligent method for assisting infant and toddler rehabilitation training by recognizing movement trajectories, such as... Figure 1 As shown, the method includes the following steps:

[0040] S101: Data Acquisition and Model Preparation.

[0041] Specifically, the process of constructing the infant keypoint-annotated image dataset includes: using high-definition cameras and depth cameras (such as Intel RealSense D435i) in various scenarios such as home environments and medical rooms to collect video data covering gross motor activities of infants aged 0-36 months, including crawling, standing with support, walking, running, and jumping. Keyframe images are extracted from these data. Using the open-source annotation tool LabelMe, pediatric rehabilitation experts accurately annotate 17 key points of the infants in the images (including 3 on the top of the head, 2 on the neck, left and right shoulders, left and right elbows, left and right wrists, left and right hips, left and right knees, and left and right ankles), forming an annotation file. A total of approximately 50,000 annotated images are randomly divided into training, validation, and test sets in a 7:2:1 ratio.

[0042] Using the aforementioned training and validation sets, transfer learning was performed on the open-source pose estimation model HRNet (High-Resolution Net) within the PyTorch framework. During training, the input images were uniformly scaled to 256×256 pixels, the Adam optimizer was used, the initial learning rate was set to 1e-3, and the model was trained for 100 epochs. The model weights that performed best on the validation set were saved to obtain the end-to-end infant human keypoint detection model.

[0043] Meanwhile, a multi-target keypoint tracker was configured using the Python programming language and libraries such as OpenCV and FilterPy. This tracker has a preset confidence threshold of 0.7 and implements data association logic based on Kalman filtering and the Hungarian algorithm for subsequent trajectory maintenance.

[0044] S102: Real-time detection and trajectory tracking.

[0045] In a home environment, mount the depth camera above the crib or to the side of the playpen to ensure complete coverage of the infant's activity area. The camera captures a real-time RGB-D video stream at 30 frames per second.

[0046] For each frame of the RGB image, histogram equalization is first performed to enhance contrast, and then the image is fed into the trained HRNet model for inference. The model outputs the coordinates (x, y) of all detected keypoints in that frame, along with their corresponding confidence scores. For example, the "left wrist" keypoint is detected at coordinates (120, 300), with a confidence score of 0.92.

[0047] Subsequently, the detection results (coordinates and confidence scores) of all keypoints in the current frame are input into the multi-target keypoint tracker. The tracker internally executes the following sub-steps:

[0048] Confidence screening: Filter out test results with a confidence level lower than 0.7.

[0049] Trajectory prediction: For an existing trajectory (e.g., the "right knee" trajectory with ID 3), a Kalman filter is used to predict its expected position in the current frame based on its historical coordinates over the past 5 frames.

[0050] Data Association: Calculate the Euclidean distance between all predicted trajectory locations and all valid detection locations in the current frame, and construct a cost matrix. Use the Hungarian algorithm to find the optimal match. If the Euclidean distance between a predicted trajectory location and a detection location is less than 20 pixels, the match is considered successful.

[0051] Trajectory Update: For a successfully matched trajectory, update its Kalman filter state using the matched detection coordinates and append the coordinates to its historical coordinate sequence.

[0052] Track Management: For existing tracks that fail to match, their "loss count" is incremented by 1; if the "loss count" exceeds 10 consecutive frames, the track is terminated and deleted. For any new unmatched detection point, a new track is created for it, a new unique ID is assigned (e.g., ID=15), and its Kalman filter is initialized.

[0053] Through the above process, a stable trajectory with a unique ID and containing time-series coordinates (such as [(t1, x1, y1), (t2, x2, y2), …]) is generated for each physical key point (such as the left wrist of an infant) throughout the video sequence.

[0054] S103: Motion trajectory generation and feature quantization.

[0055] Taking the evaluation of the "crawling" action as an example, four key point trajectories were selected: "left wrist", "right wrist", "left knee joint", and "right knee joint". For the "left wrist" trajectory (ID=5), the two-dimensional coordinates (x, y) of all its frames were connected to form the two-dimensional motion trajectory of the left wrist on the image plane.

[0056] Feature quantization of the trajectory:

[0057] Path length: Calculate and sum the Euclidean distances between adjacent frame coordinates.

[0058] Average speed: Path length divided by total trajectory time.

[0059] Motion smoothness: Calculates the rate of change of the curvature of the trajectory.

[0060] Joint angles: For example, calculate the angle formed by "shoulder-elbow-wrist" in each frame and count its range of variation.

[0061] After extracting the above features, they are combined to form a multi-dimensional quantitative feature vector, which is used to characterize the movement pattern of the left wrist during this crawling action.

[0062] If a depth camera is used, RGB-D data can be used to back-project the two-dimensional pixel coordinates (x, y) and depth value d into three-dimensional spatial coordinates (X, Y, Z) through camera intrinsic parameters, thereby constructing a three-dimensional motion trajectory and calculating features such as velocity and acceleration in three-dimensional space.

[0063] S104: Motor Analysis and Developmental Assessment.

[0064] Assume the current assessment subject is a 12-month-old infant. The system retrieves the feature vector template of the standard "crawling" movement for 12-month-old infants from the standard infant gross motor development milestone database. This template contains typical statistical ranges (such as mean and standard deviation) for limb movement path length, speed, and joint range of motion when infants of this age crawl.

[0065] The current infant crawling feature vector obtained in step S103 is compared with the standard template. A motion analysis model (such as a pre-trained Support Vector Machine (SVM) classifier or by calculating cosine similarity) scores each feature. For example, the Z-score of the average path length of the current left wrist and the standard path length is calculated. If the Z-score is within the range [-1, 1], the feature is considered satisfactory; otherwise, it is considered biased. A motion assessment report is generated by combining the scores of all feature items, including: "Overall motion completion: 85%; left hand propulsion is slightly weak (speed is low); right knee coordination is good; it is recommended to strengthen left upper limb strength training."

[0066] S105: Personalized guidance generation and interactive feedback.

[0067] According to the aforementioned evaluation report, the personalized recommendation engine retrieves training programs for "strengthening the left upper limb strength" from the training action library, such as the "left-side single-arm support object retrieval game," and generates detailed instruction text and demonstration animations.

[0068] On the parent's mobile app (user interaction terminal), the interface simultaneously displays the following:

[0069] In the live video feed, the infant's current left wrist position is marked with a red highlight circle.

[0070] On one side of the screen, a small window plays an animation of standard crawling movements, which is then transparently overlaid and contrasted with the real-time movements of the infant.

[0071] The voice prompt reads: "Baby, you can push harder with your left hand!"

[0072] On the assessment report page, a line graph showing the developmental trend of "crawling ability" from 10 months to the current 12 months of age is generated.

[0073] Parents can follow the guidance to conduct interactive training with infants and toddlers.

[0074] Example 2:

[0075] This embodiment provides an intelligent motion trajectory recognition auxiliary system for infant rehabilitation training, which implements the method of Embodiment 1. Its architecture is as follows: Figure 2 As shown.

[0076] The system includes:

[0077] Data acquisition module: This is a hardware unit integrated into the head of the smart crib, with a built-in RGB-D camera (such as Orbbec Astra Pro) and microphone array, used to collect audio and video data and depth information of infants' activities.

[0078] Edge computing unit: Utilizing a Raspberry Pi 4B computing module, it deploys a lightweight keypoint detection model accelerated by TensorRT (such as MobileNetV2, a lightweight version of HRNet for the backbone). This unit is responsible for decoding, preprocessing, and performing preliminary keypoint detection on the raw video stream, and then compressing the detection results (keypoint coordinates and confidence scores) before uploading them via Wi-Fi.

[0079] Cloud-based data processing and analysis platform: Deployed on Alibaba Cloud ECS servers, using a microservice architecture. Among its components:

[0080] The trajectory tracking service (a standalone Docker container) continuously runs the multi-target keypoint tracker, receives the detection result stream from the edge computing unit, outputs time-series trajectory data with IDs, and publishes the data to a Redis message queue.

[0081] The feature assessment service subscribes to trajectory data, performs feature quantification and comparative analysis with a standard database of infant gross motor development milestones stored in a cloud-based MySQL database, and generates an assessment report.

[0082] The recommendation service queries and generates guidance plans from the training action library, which is also stored in MySQL, based on the evaluation report.

[0083] The database update mechanism is as follows: a weekly scheduled task will anonymize and aggregate the feature vectors of all users, perform statistical analysis, and if a significant change is found in the distribution of group data for a certain action pattern (such as "13-month-old independent walking gait"), the standard template in the database will be iteratively updated after expert review.

[0084] The user interaction terminal is primarily the "Baby Growth Assistant" app (Android / iOS) on parents' mobile phones. The app communicates with cloud services via an HTTPS persistent connection, receives assessment reports and guidance plans, and displays them in rich media format. Simultaneously, the app provides interactive functions such as historical data review, training plan settings, and online expert consultation.

[0085] Example 3:

[0086] This embodiment provides a computer-readable storage medium (such as a USB flash drive or a virtual disk in cloud storage) on which a computer program is stored. When the program is executed by the processor of a computer device (such as the processor built into a home smart camera or a home gateway), the processor can load and run the program instructions, thereby implementing all or part of the method flow described in steps S102 to S105 of Embodiment 1, for example, completing the entire process from receiving a video stream to generating screen guidance information.

[0087] This embodiment also provides a computer device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the method described in Embodiment 1. This computer device can be a dedicated device integrated into the smart crib head, or it can be a smart TV or tablet computer with strong computing capabilities configured in the home.

[0088] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0089] The accompanying drawings of the embodiments disclosed in this invention only involve structures relevant to the embodiments disclosed in this invention. Other structures can be referred to with common designs. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other.

[0090] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for assisting infant and toddler rehabilitation training by intelligently recognizing movement trajectories, characterized in that, Includes the following steps: S1. Data Acquisition and Model Preparation: Construct an image dataset of keypoint-annotated images of infants and toddlers, and use this dataset to train an end-to-end human keypoint detection model for infants and toddlers; at the same time, configure a multi-target keypoint tracker based on temporal prediction and data association. S2. Real-time detection and trajectory tracking: Acquire real-time video streams of the infant's activity area through visual acquisition devices; For each frame of image, the trained keypoint detection model is used to detect keypoints and output the coordinates and confidence scores of keypoints including at least the head, shoulder, elbow, hip, knee, and ankle joints. Subsequently, the key point detection results are input into the multi-target key point tracker, and a unique trajectory ID is assigned to and maintained for each detected key point through cross-frame data association, generating continuous and stable time series key point coordinate data; S3. Motion trajectory generation and feature quantization: Based on the coordinate data of key points with the same trajectory ID, construct the motion trajectory of a specific body part or body center of gravity in two-dimensional or three-dimensional space; perform parameterized calculation on the motion trajectory, and extract one or more quantified features including trajectory path length, average motion speed, peak acceleration, motion smoothness, joint angle change range and motion cycle to form a quantized feature vector representing the current motion; S4. Motion Analysis and Developmental Assessment: The quantitative feature vector is input into a preset motion analysis model and compared with the standard motion template features of the corresponding age in the standard infant gross motor development milestone database; the motion analysis model comprehensively evaluates the completion, standardization and deviation from the developmental standard of the current motion by calculating feature similarity or distance, and generates a motion assessment report containing quantitative scores and qualitative descriptions; S5. Personalized guidance generation and interactive feedback: Based on the action assessment report, a targeted follow-up training guidance plan is generated by adaptive matching or combination from the preset training action library; through the graphical user interface, the assessment results and training guidance are fed back to the caregiver in real time in one or more ways, such as overlaying comparison animation, highlighting deviation areas, voice prompts and development progress charts.

2. The method according to claim 1, characterized in that, In step S2, the workflow of the multi-target keypoint tracker includes: S21. Confidence Filtering: Set a confidence threshold and only send key point detection results with a confidence level higher than the threshold into the tracking process; S22. Trajectory Initialization and Management: Assign a new trajectory ID to the first valid key point that appears in each frame, and establish a trajectory dictionary to store its historical coordinate sequence using the trajectory ID as the key; S23. State prediction and data association: For an existing trajectory, use a Kalman filter to predict the expected position in the current frame based on its historical coordinates; calculate the correlation between the predicted position and the actual detected key point position in the current frame, use the Hungarian algorithm for optimal matching, and update the coordinates of the corresponding trajectory if a successful match is achieved. S24. Trajectory Lifecycle Management: Retain the count of existing trajectories that fail to match. If the count is not updated after a preset number of frames, the trajectory is determined to be terminated and deleted. For new detection points that do not match, initialize them as new trajectories.

3. The method according to claim 1 or 2, characterized in that, In step S1, constructing the infant key point labeled image dataset includes: collecting images of infant activities at different ages, in different scenes, and under different lighting conditions; using annotation tools to accurately label the key points of the infant's body in the images; and dividing the dataset into training set, validation set, and test set according to the proportions for model training and evaluation.

4. The method according to claim 1, characterized in that, In step S3, if the visual acquisition device is a depth camera or an RGB-D camera, the depth information is used to calculate the three-dimensional spatial coordinates of the key points, construct the three-dimensional motion trajectory, and perform feature quantization in the three-dimensional space.

5. The method according to claim 1, characterized in that, In step S4, the standard infant gross motor development milestone database is dynamically updated, and its update data comes from the statistical analysis results of trajectory feature vectors after anonymization of a large number of infant users.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 5.

7. An infant rehabilitation training assistance system for intelligently recognizing movement trajectories to implement the method of any one of claims 1-5, characterized in that, include: The data acquisition module, deployed in the infant and toddler activity area, includes at least one visual acquisition device for continuously acquiring video stream data; An edge computing unit, connected to the data acquisition module, is used to run the lightweight key point detection model and perform preliminary detection and preprocessing. A cloud-based data processing and analysis platform, communicatively connected to the edge computing unit, includes: Trajectory tracking server: used to run the multi-target keypoint tracker, receive keypoint detection results and output time-series trajectory data; Feature extraction and evaluation engine: used to perform motion trajectory generation, feature quantization, and comparative analysis with standard databases to generate motion evaluation reports; Personalized recommendation engine: It has the built-in training action library and generates personalized training guidance plans based on the evaluation report; The user interaction terminal is connected to the cloud data processing and analysis platform to receive and visualize the action evaluation report and training guidance plan, and to provide a human-computer interaction interface.

8. The system according to claim 7, characterized in that, The visual acquisition device integrates an infrared transmitter and a receiver to form a depth sensing module, which is used to acquire depth image information of infants and young children.

9. The system according to claim 7, characterized in that, The system also includes a parent-side mobile application, which serves as one implementation of the user interaction terminal. The parent-side mobile application provides functions such as video viewing, historical evaluation reports, training plan push notifications, and community communication.

10. The system according to claim 7, characterized in that, The cloud-based data processing and analysis platform adopts a microservice architecture, with the trajectory tracking server, feature extraction and evaluation engine, and personalized recommendation engine deployed as independent services and exchanging data through message queues.

Images