Dance treatment posture correction method based on deep learning

Abstract

The invention relates to the technical field of dance treatment and intelligent rehabilitation, in particular to a dance treatment posture correction method based on deep learning. According to the method, a multi-modal sensing system is constructed, three-dimensional attitude data of a dancer is acquired by using an RGB-D visual sensor, an inertial measurement unit (IMU) and motion capture equipment, a depth attitude recognition model is established in combination with a convolutional neural network (CNN) and a time sequence long-short-term memory network (LSTM), and feature extraction and time sequence analysis are performed on dancing motions. The system evaluates the difference between the posture of the dancer and the standard dance posture in real time through the posture error calculation module, outputs vision and voice dual-mode feedback according to a correction algorithm, guides the dancer to adjust the body posture, and achieves personalized dance treatment and exercise rehabilitation. According to the method, high-precision posture correction and situation feedback can be realized under different dance types and treatment scenes, and the scientificity and the intelligent level of dance treatment are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of dance therapy and intelligent rehabilitation technology, aiming to solve problems existing in current dance therapy processes, such as guidance relying on human experience, strong subjectivity in posture assessment, lack of real-time feedback mechanisms, and lack of individual adaptability in movement correction. In existing technologies, posture correction often relies solely on video comparison or verbal prompts from experts, making it difficult to achieve quantitative error calculation and adaptive feedback adjustment, resulting in insufficient scientific rigor and precision in dance therapy for rehabilitation, psychological intervention, and limb coordination training. Background Technology

[0002] Dance / Movement Therapy (DMT), a non-pharmacological therapy that combines movement, rhythm, and psychological intervention, has gained significant attention in rehabilitation and psychological intervention in recent years. Traditional dance therapy heavily relies on the therapist's visual observation and experience, resulting in strong subjectivity, difficulty in quantification, delayed feedback, and challenges in large-scale concurrent training. Modern motion capture (such as IMU + visual skeleton recognition) and physiological signal monitoring (heart rate, GSR, surface electromyography, EMG) offer the possibility of establishing objective quantitative assessments; however, existing systems often process posture or physiological data independently, lacking a holistic approach that couples these two at the temporal and semantic levels to achieve real-time adaptive feedback. Furthermore, the high degree of freedom and individual variability of complex dance movements place higher demands on the system's generalization ability and feedback safety. Therefore, there is an urgent need to propose a closed-loop method that integrates deep learning posture assessment, physiological state perception, and personalized feedback generation to achieve synergistic optimization of "movement-physiology-feedback," thereby improving the accuracy, timeliness, and adaptability of dance therapy.

[0003] Traditional dance posture correction systems typically rely solely on visual recognition or angular changes from inertial sensors to determine the correctness of movements, failing to incorporate physiological load and emotional state into their analysis. For example, in therapeutic dance training, muscle tension or elevated heart rate can lead to stiffness or abnormal amplitude in movements, but the system cannot distinguish between technical deviations and physiological responses. Therefore, simple "posture deviation warnings" often cause psychological stress for patients and can even reduce rehabilitation effectiveness. To address this, this invention proposes a novel posture-physiology fusion adaptive correction mechanism: by simultaneously acquiring human kinematic signals (IMU and skeletal point coordinates) and physiological signals (heart rate, skin conductance, electromyography), a deep learning model is used to analyze movement deviations at both temporal and semantic levels, while simultaneously calculating physiological state weights as feedback control parameters. This method not only identifies the spatial location and amplitude of posture errors but also dynamically adjusts the feedback intensity and rhythm based on the individual's current physical state, preventing "overcorrection" or "overstimulation." Summary of the Invention

[0004] A deep learning-based method for posture correction in dance therapy, characterized by the following steps: Dance movement data acquisition and preprocessing, including: The system utilizes multimodal sensing devices to acquire the dancer's motion data, including 2D image sequences and depth images captured by an RGB-D camera; IMU sensors measuring the acceleration and angular velocity of various body parts in real time; and pressure sensing pads detecting plantar pressure and center of gravity shift. Data from all sensors is fused using a unified timestamp synchronization mechanism to form a multi-source time-series dataset. To improve model recognition accuracy, the following preprocessing operations are performed on the input data: Three-dimensional median filtering and Gaussian smoothing are used to eliminate noise; the attitude coordinates are normalized. , Data augmentation algorithms (mirror rotation, temporal perturbation, and noise injection) are used to expand the sample set to enhance the model's generalization performance.

[0005] Preferably, the system extracts key postures from standard movement samples collected by professional dance therapists and rehabilitation experts to establish a standard dance posture database. The system then extracts the coordinates of key human points in each frame using a human skeleton detection algorithm (such as OpenPose or HRNet) to form a standard three-dimensional posture matrix. , in This represents the number of joints. The system further calculates the temporal smoothing weights of the motion sequence and performs a dynamic weighted average on the standard samples to generate a standard trajectory template for each dance move: , in The weights of the expert samples satisfy the following conditions: This database can establish classification indexes for different dance styles (such as ballet, modern dance, folk dance, etc.) and different therapeutic purposes (coordination, balance, psychological relaxation), which can be used for subsequent matching and identification.

[0006] Preferably, for pose recognition and deep feature extraction, the dancer's real-time pose sequence is input into a deep learning network for recognition and feature extraction. The network structure adopts a joint CNN and LSTM model, with the front-end convolutional layer used for spatial feature extraction and the back-end LSTM layer used for temporal dependency modeling. The network input is a set of temporally continuous pose frames. The output is the corresponding action category and pose feature vector F(t). The network loss function is defined as: , in, This is the weight decay coefficient. Let be the parameter norm. The network is iteratively trained using backpropagation and the Adam optimization algorithm until the loss function converges. A transfer learning strategy is introduced during model training to reduce data requirements and improve adaptability among different dancers.

[0007] Preferably, the system calculates and quantifies posture errors by using a model to calculate the deviation between the dancer's current posture and the standard posture in real time as the dancer performs dance movements. Error assessment indicators include spatial error, angular error, and center of gravity deviation.

[0008] Spatial error: , Angular error: For adjacent joints a, b, c, calculate their included angle difference: , Overall attitude quality score: , The two endpoints of the vibration threshold range are respectively denoted as the lower vibration threshold and the upper vibration threshold. in This represents the balance stability score. , , This is the weighting coefficient. Using this indicator, the system can quantitatively assess how closely a dancer's posture resembles a standard dance posture.

[0009] Preferably, posture correction and multimodal feedback are used when the posture quality score is... Below the preset threshold At this time, the system automatically triggers the corrective feedback mechanism. The feedback module includes three modes: visual enhancement feedback, voice guidance feedback, and motion trajectory guidance. Visual Enhancement Feedback: The system displays the areas of difference between the dancer and the standard movements in real time on the screen or projected image in the form of a virtual skeleton, and uses color gradients to indicate the size of the error.

[0010] Voice guidance feedback: The system generates personalized voice prompts based on the direction of the error, such as "raise the left shoulder" or "shift the center of gravity to the left," to help dancers make timely adjustments.

[0011] Movement trajectory guidance: The corresponding joint trajectory is projected onto the screen as a virtual arrow, guiding the dancer to complete the corrective movement along the correction path.

[0012] The feedback system features a low latency of 0.2 seconds, enabling near real-time human-computer interaction.

[0013] The preferred adaptive learning and model update mechanism allows the system to automatically record the movement history and corrective effects of each dancer during dance therapy training. When the error distribution tends to stabilize or deviates from the standard movement curve after multiple training sessions, the system automatically triggers the model retraining module.

[0014] Sample accumulation: The system stores pose frames with large errors as "abnormal samples" into an adaptive sample library; Weight update: Adjust network parameter weights by combining historical training results and treatment feedback; Personalized model generation: obtaining individualized models through differentiated training. Its loss function is: , in This represents the individual bias coefficient. Through periodic optimization, the system can continuously improve recognition accuracy and individual adaptability, achieving dynamic dance therapy assistance.

[0015] Preferably, the system of the present invention includes the following main modules: Posture acquisition module (101): includes an RGB-D camera, an IMU sensor and a pressure detection pad, used to acquire dancer posture information, center of gravity movement and limb angle changes.

[0016] Feature extraction module (102): Embedded CNN-LSTM deep neural network to realize spatiotemporal feature learning and action sequence recognition.

[0017] Error assessment module (103): Calculates posture deviation, joint angle difference, and posture score. This generates real-time quantitative results.

[0018] Feedback and interaction module (104): Outputs voice prompts, virtual images and trajectory guidance to dancers to achieve real-time guidance and multimodal interaction.

[0019] Model optimization module (105): Updates deep network parameters through an online learning mechanism to form a personalized training model.

[0020] Compared with the prior art, the beneficial effects of the present invention are: Dancers connect to a deep analysis server via motion capture devices. The server's analysis results are transmitted to a feedback terminal via a communication module, forming a closed-loop control system. Multimodal data fusion and deep temporal networks are introduced into pose recognition to achieve millisecond-level data alignment and high-precision pose prediction. This is achieved through a pose quality index. The dynamic error curve allows therapists to quantitatively assess dancers' movement execution. The human-computer interaction mode, combining visual, audio, and projection elements, significantly improves dancers' learning efficiency and therapeutic immersion. Through a continuous learning mechanism, the system can self-adjust to different ages, rehabilitation stages, and dance styles, forming personalized dance posture correction plans. This invention is not only applicable to dance therapy but can also be extended to fields such as geriatric rehabilitation, posture training, sports rehabilitation, and posture assessment, possessing broad industrial application prospects.

[0021] The technical framework of this invention has high versatility. Its core innovation lies in adopting a closed-loop system architecture of multi-sensor fusion + deep learning recognition + real-time feedback correction; and proposing a posture quality index. The invention employs a quantitative model for real-time assessment of posture quality; combined with a cloud-based data sharing mechanism, it enables collaborative training and experience transfer across different user models; a cognitive behavioral science-based movement deviation labeling mechanism is introduced during model training to ensure the system's feedback aligns with the principles of dance therapy psychology; and it supports distributed deployment, making it applicable to dance studios, rehabilitation centers, or home therapy settings. Through these technical designs, the invention achieves intelligent assessment, real-time feedback, and personalized guidance in dance therapy, providing a quantifiable, learnable, and feedback-enabled technological foundation for rehabilitation training and psychotherapy. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the overall structure of a deep learning-based dance therapy posture correction system. Figure 2 Flowchart for dance posture recognition and correction processing; Figure 3 This is a schematic diagram of a deep learning neural network structure; Figure 4 This diagram illustrates the relationship between attitude error calculation and feedback mechanism. Figure 5 A schematic diagram illustrating an application scenario for a dance therapy system; Detailed Implementation

[0023] To make the technical solution of the present invention clearer and more understandable, the present invention will be further described below with reference to the accompanying drawings and specific embodiments. These embodiments are merely illustrative examples and are not intended to limit the scope of protection of the present invention.

[0024] like Figure 1 As shown, the deep learning-based dance therapy posture correction system of the present invention mainly includes: The module includes a posture acquisition module 101, a feature analysis module 102, an error evaluation module 103, a feedback interaction module 104, and a model optimization module 105.

[0025] The posture acquisition module 101 is responsible for collecting real-time motion information from the dancer, including RGB-D images, inertial measurement data, and plantar pressure distribution; the feature analysis module 102 embeds a deep neural network structure to extract posture features and predict movement categories; and the error evaluation module 103 calculates the multidimensional deviation between the dancer's current movement and the standard dance posture to form a posture quality index. The feedback interaction module 104 outputs guidance information through visual overlay, voice prompts, and virtual trajectories; the model optimization module 105 retrains the network by combining the dancer's training history data to generate an individualized correction model. The entire system operates collaboratively with the cloud training platform through local edge computing nodes and can be deployed and used in different venues (dance studios, rehabilitation centers, or home environments).

[0026] Example 1: See Figure 2 The system first collects dancer movement data through multiple sensors. An RGB-D camera captures two-dimensional images and depth information of the human skeleton, an IMU module detects linear acceleration and angular velocity at key points, and a pressure-sensing pad records the force distribution on the dancer's feet and the trajectory of changes in the center of gravity. The system uses a unified timestamp synchronization protocol to align the data output from each sensor with millisecond-level precision. The collected data stream undergoes filtering and normalization operations by the local processor, using the following formula: , in , Here, represents the mean and standard deviation of the data in the i-th channel, respectively. After data synchronization, the system generates a high-dimensional input matrix consisting of time series data. ,

[0027] in RGB-D image frame For IMU signal vectors, This is the pressure distribution matrix. The standard dance posture database consists of movement samples provided by multiple professional dance therapists. The system identifies key nodes (head, shoulder, elbow, knee, ankle, etc.) in each frame through a human skeleton extraction model and represents their three-dimensional coordinates as follows: , Standard trajectory is calculated using a multi-sample fusion method: , Simultaneously, the trajectory is smoothed temporally and redundant frames are compressed to ensure the temporal continuity and spatial accuracy of standard dance postures. Each standard dance posture template is accompanied by metadata, including dance style, training objectives (such as balance, coordination, and psychological relaxation), and target audience, for subsequent personalized matching.

[0028] Example 2: See Figure 3 The core of the feature analysis module is a hybrid CNN-LSTM deep network. The convolutional network part is used to extract spatial features, with image frames as input. Skeleton heatmap features are obtained through multi-layer convolution, pooling, and batch normalization operations. , , The temporal network portion (LSTM) is used to capture dynamic changes in the time dimension, transforming the feature sequence... Input LSTM unit: , The output is action timing features. The network then generates pose prediction results. The overall loss function of the network is: , Training employs the Adam optimizer with adaptive learning rate decay, and the model is trained on a GPU for accelerated performance. Using this model, the system achieves highly robust human pose recognition and action classification in complex backgrounds and under varying lighting conditions.

[0029] Example 3: Attitude Error Calculation and Feedback Strategy (see [reference]) Figure 4 The error assessment module calculates the multidimensional deviation between the dancer's current posture and the standard dance posture based on the deep learning recognition results. The error calculation consists of three parts: Spatial deviation : , Angle deviation The joint rotation difference is calculated using the angle between joint vectors: , Center of gravity deviation The center of gravity shift is calculated using the plantar pressure matrix: , The final attitude quality index is defined as: , when The system automatically triggers a feedback mechanism at that time. The feedback interaction module adopts a three-layer mechanism: The visual feedback layer overlays skeletal images of dancers and standard poses onto a display screen or virtual projection, using color to distinguish deviations; The voice feedback layer outputs natural voice prompts based on the wrong direction, such as "relax your left shoulder" and "raise your right arm". The motion trajectory guidance layer dynamically displays the correction direction as arrows in the virtual interface. To ensure a natural interactive experience, the system controls the visual latency to within 200 milliseconds and automatically adjusts the prompt frequency based on the dancer's reaction time.

[0030] Example 4: See Figure 4 The model optimization module 105 receives feedback from the error evaluation module and retrains the network parameters based on individual performance. The system employs a transfer learning strategy, using the group model parameters as initial weights and then fine-tuning them using individual dancer data. The adaptive learning algorithm is defined as follows: , in For learning rate, This is the regularization coefficient. Through continuous iteration, the system can gradually form a personalized correction model that matches the dancer's movement characteristics, achieving cross-individual transfer and long-term self-learning. The model update cycle can be automatically adjusted according to the user's training frequency. When the training error continues to decrease or stabilizes within the threshold range, the system pauses retraining to reduce computational resource consumption.

[0031] Example 5: See Figure 5 The system of the present invention can be widely applied in the following scenarios: Rehabilitation centers assist patients with stroke, Parkinson's disease, and other conditions with limb coordination training, improving rehabilitation efficiency through postural feedback. Dance teaching centers utilize a standard dance posture database to achieve intelligent assessment and precise teaching guidance. Home dance training allows users to practice independently at home using consumer-grade cameras and terminal devices; the system corrects movements in real time and provides psychological encouragement. Psychological therapy support combines dance with physical movement, along with visual feedback and voice guidance, to help users alleviate anxiety and depression. The system can run on a local area network or in the cloud and supports 5G remote connectivity. Therapists can monitor trainees' training status through a management platform, viewing posture scoring curves, number of corrections, and progress trends.

[0032] The results show that the system of this invention exhibits significant accuracy in posture recognition and real-time feedback in multiple dance therapy experiments. With a standard dance posture database containing 2000 samples, the posture recognition accuracy reached 97.3%, the average feedback delay was 0.18 seconds, and the average dancer posture error was reduced by 43.6%. The closed-loop learning and feedback mechanism of this invention makes the dance therapy process more scientific, quantifiable, and personalized, and has significant promotional value in the fields of rehabilitation therapy and arts education.

Claims

1. A deep learning-based method for posture correction in dance therapy, characterized in that, Includes the following steps: Dance data acquisition and synchronization utilizes multimodal sensing devices to collect dancers' real-time motion data, including RGB-D images, inertial measurement unit (IMU) signals, and plantar pressure distribution data. A unified timestamp synchronization mechanism is used to align multi-source data and generate multi-dimensional motion sequences of dancers. Standard dance posture modeling: A standard dance posture database is established based on professional dance therapy samples. The three-dimensional coordinates of key points are extracted using a human skeleton recognition algorithm to form a standard posture matrix. And generate a standard motion trajectory template using a weighted average; Deep learning pose recognition inputs dancers' pose data into a deep learning model composed of a convolutional neural network (CNN) and a long short-term memory network (LSTM), extracting spatial and temporal features to obtain the dancer's pose vector for the current movement. ; Posture error calculation involves determining spatial error, angular error, and center of gravity deviation based on the joint offset between the dancer's posture and the standard dance posture, to obtain the posture quality index. : ， in, , , These are weighting coefficients; Real-time posture correction and feedback: When the posture quality index Q(t) is lower than the set threshold, the system triggers a feedback mechanism, outputting movement correction information to the dancer through visual overlay, voice prompts and virtual trajectory guidance to achieve posture correction. The system uses adaptive optimization to update the parameters of the deep learning model based on the dancer's historical training data and uses transfer learning mechanism to form an individualized model, thereby achieving personalized dance therapy posture correction.

2. The method according to claim 1, characterized in that, In the dance data acquisition step, the RGB-D image frames, IMU signals, and pressure matrix are all accompanied by timestamp information, and frame-level alignment is achieved through an interpolation synchronization algorithm.

3. The method according to claim 1, characterized in that, The standard dance posture database is divided into sub-databases based on dance type, such as ballet, modern dance, folk dance, and rehabilitation dance. Each sub-database contains diverse standard posture samples and metadata tags.

4. The method according to claim 1, characterized in that, The convolutional layers of the deep learning model are used to extract spatial skeleton features from the dancer's image, and the LSTM layers are used to capture temporal features of the movements. The loss function of the network is defined as: ， in This is the regularization parameter.

5. The method according to claim 1, characterized in that, The attitude error calculation module calculates the following indicators: Spatial error This reflects the spatial distance deviation of key points; Angular error This reflects the difference in joint rotation angle; Center of gravity deviation This reflects the stability of the body's center of gravity; The system combines the three factors to obtain an attitude score.

6. The method according to claim 1, characterized in that, The feedback mechanism includes three modes: visual feedback, voice feedback, and trajectory guidance. Visual feedback uses color gradients to display the error magnitude, voice feedback outputs correction instructions in natural language, and trajectory guidance displays the correction direction through virtual arrows.

7. The method according to claim 6, characterized in that, In the model adaptive optimization step, the system retrains the weights of the deep learning network based on error analysis results to form an individualized model. Its loss function is: 。 8. The method according to claim 1, characterized in that, The system adopts an edge computing architecture, where dance posture recognition is completed on the local terminal, and model optimization and storage are executed in the cloud to reduce latency and support multi-user concurrency.

9. The method according to claim 1, characterized in that, When persistent deviations in a dancer's posture are detected, the system optimizes the user experience by dynamically adjusting the feedback frequency and intensity of prompts. A deep learning-based dance therapy posture correction system, characterized in that it includes: Posture acquisition module (101): used to acquire the dancer's motion images, inertial signals and pressure data, and generate multi-source synchronous data streams; Feature analysis module (102): used to execute the CNN-LSTM fusion network model, extract spatial and temporal features, and output the dancer's motion feature vector; Error assessment module (103): Used to calculate the deviation between the dancer's posture and the standard dance posture, and generate a posture quality index. ; Feedback interaction module (104): used to output visual, voice and motion trajectory feedback based on posture quality index to realize dancer movement correction; Model optimization module (105): used to update the parameters of the deep learning model based on the dancer's training history data to achieve individualized correction.

10. The method according to claims 1 to 9, characterized in that, The system also includes a posture data visualization module, which is used to draw posture scoring curves, error distribution maps, and training progress reports to assist therapists in making assessments.

Images