A meditation posture adjustment method based on a light neural network and an acoustic sense

Abstract

The application discloses a meditation posture adjustment method based on a light neural network and an acoustic sense, and has the characteristics that the method specifically comprises the following steps: collecting real-time video data in a user's meditation process, detecting and extracting features of human body skeleton key points of an image by using a light neural network model, analyzing key point position information to calculate the deviation of a current posture from a standard meditation posture, matching corresponding voice action instructions according to the deviation type, playing voice guidance to guide the user to adjust the body action in real time, and cyclically monitoring the posture change and dynamically updating the voice feedback until the posture is corrected. The application effectively reduces the calculation complexity of the model to adapt to mobile terminal deployment, and innovatively uses voice instructions to replace visual feedback, solves the pain point that the user needs to open eyes to confirm the posture in the meditation process and destroys the immersion, realizes real-time and accurate posture correction in a closed eye state, and has good application prospect.

Description

Technical Field

[0001] This invention relates to the field of computer vision and intelligent health interaction technology, specifically to a meditation posture adjustment method based on lightweight neural networks and acoustics. Background Technology

[0002] With the fast pace of modern life, people face increasing psychological pressure. Meditation, as an effective way to regulate the mind and body and relieve stress, has received widespread attention. Maintaining correct posture (such as a straight spine and relaxed shoulders) is crucial for entering a meditative state and avoiding physical injury during meditation. However, for beginners, it is difficult to maintain the correct posture for an extended period without the guidance of a professional instructor.

[0003] Early assistance methods primarily relied on in-person guidance from professional tutors, but these were costly and limited by time and space. With the development of artificial intelligence technology, a series of computer vision-based pose recognition applications have emerged. Recently, deep learning-based human pose estimation methods have achieved excellent results, utilizing convolutional neural networks to locate key points on the human body and accurately capture user poses.

[0004] However, in practical meditation applications, existing technical solutions have the following problems: First, mainstream high-precision pose estimation algorithms have a large number of parameters and high computational complexity, making them difficult to run smoothly on mobile devices such as smartphones and tablets. They often rely on cloud computing, resulting in high latency and privacy risks. Second, and more critically, existing pose guidance systems mostly use visual feedback mechanisms (such as displaying skeletal overlay maps or text prompts on the screen). This requires users to keep their eyes open and observe the screen during meditation. This visual interaction directly disrupts the immersive state of closed-eye meditation, interrupting the user's focus and significantly reducing the effectiveness of meditation. Furthermore, simple threshold judgments often ignore the continuity of human movements and jitter errors, easily leading to false alarms. Therefore, traditional guidance methods that rely primarily on visual feedback and high-computational models cannot effectively address this problem. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing a meditation posture adjustment method based on lightweight neural networks and acoustics. It employs a lightweight neural network to achieve low-latency posture capture on mobile devices, constructing a non-visual interactive feedback mechanism that simulates the voice guidance process of a professional instructor. Intelligent cropping and center point regression techniques are used to extract key points of the human skeleton, calculate the geometric angles and positional relationships between these key points, and convert posture deviations into specific voice commands. Real-time monitoring, deviation analysis, voice matching, and dynamic feedback are then performed sequentially. Finally, acoustic enhancement guides the user to adjust their movements, completing posture correction without requiring the user to open their eyes. This invention can effectively run on resource-constrained devices, creatively solving the problem of visual feedback disrupting the immersion of meditation. By using voice commands instead of screen displays, it achieves accurate posture correction even with eyes closed. The method is simple, effective, and has promising application prospects.

[0006] The objective of this invention is achieved as follows: A meditation posture adjustment method based on lightweight neural networks and acoustics, comprising the following steps: Step 1: Collect real-time video stream data during the user's meditation process and perform image preprocessing to obtain a standardized input image sequence; Step 2: Use a lightweight convolutional neural network model to extract features from the input image sequence, detect and output the coordinate information of key points of the human skeleton; Step 3: Construct the geometric feature vector of the current pose based on the coordinates of the skeletal key points, and calculate the angles and relative positions of the key parts; Step 4: Compare the geometric feature vector of the current posture with the preset standard meditation posture library, and calculate the posture deviation value and deviation type; Step 5: Match the corresponding voice action command based on the posture deviation type, and generate guiding audio using the acoustic enhancement module; Step 6: Play the voice guidance instructions in real time and repeat steps 1-5 to dynamically monitor posture changes until the posture deviation value is lower than the preset threshold, thus completing the posture adjustment.

[0007] Furthermore, step 1 specifically includes: 1-1: Obtain raw video frames of the meditation scene using an infrared or RGB camera, and use a face detection algorithm to locate the user area and crop it; 1-2: Adjust the cropped image to the 192*192 resolution required by the lightweight neural network, and perform normalization to obtain a standardized input image sequence.

[0008] Furthermore, step 2 specifically includes: 2-1: Construct a lightweight backbone network containing depthwise separable convolutional layers to extract features from the input image sequence layer by layer, reducing the number of parameters and computational cost; 2-2: The shallow high-resolution features and deep strong semantic features are fused together, and the channel attention mechanism is used to enhance the feature response of key parts. Then, the coordinate information of the skeletal key points is obtained by directly regressing the key point heatmap using a differentiable Soft-Argmax operation.

[0009] Furthermore, step 2-2 specifically includes: 2-2-1: Channel concatenation is performed on shallow high-resolution features and deep strong semantic features. The channel attention mechanism is used to recalibrate the weights of the feature channels to enhance the feature response to the extremities. 2-2-2: Generate a heatmap of human key points based on the enhanced features, and use a differentiable Soft-Argmax operation to directly regress the pixel-level coordinates of key points from the heatmap to obtain the coordinate information of skeletal key points; Furthermore, step 4 specifically includes: 4-1: Select the spine, shoulders, neck, and hips as core key points to construct the core skeletal vector of the meditation posture; 4-2: Calculate the angle between adjacent limb segments using the core skeleton vector, and analyze the posture stability in conjunction with continuous frames to obtain persistent erroneous postures, and calculate the difference between the angle between each limb segment and the corresponding angle of the standard meditation posture. 4-3: Combine the differences of multiple key parts, calculate the overall posture deviation value by weighting, and determine the body part with the largest deviation to obtain the posture deviation value and deviation type.

[0010] Furthermore, step 4-2 specifically includes: 4-2-1: Calculate the actual included angle between each adjacent limb segment in the core skeleton vector; 4-2-2: Compare the calculated angle with the corresponding angle of the standard meditation posture to obtain the angle difference; 4-2-3: Combine continuous frame analysis to determine attitude stability and identify persistent erroneous attitudes.

[0011] This invention first collects real-time video stream data during the user's meditation process and performs image preprocessing. To adapt to mobile computing, a lightweight neural network model is used to extract features from the image, detect and output the coordinate information of key points of the human skeleton. Next, a geometric feature vector of the current posture is constructed based on the coordinates of the key points, the angles and relative positions of key parts are calculated, and the current posture is compared with a standard meditation posture library. Then, based on the calculated posture deviation value and deviation type, the corresponding voice action command is matched. Finally, the voice guidance command is played in real time, and the posture changes are monitored in a loop and the audio output is dynamically adjusted until the posture is corrected.

[0012] Compared with the prior art, the present invention has the following beneficial technical effects and significant technical progress: 1) This invention proposes to use a lightweight neural network for feature extraction, which enables high-precision pose estimation to run directly on mobile devices in real time without relying on expensive server resources, thus protecting user privacy and lowering the barrier to entry.

[0013] 2) The acoustic enhancement guidance mechanism proposed in this invention innovatively uses voice commands to replace traditional visual feedback. During meditation, users do not need to open their eyes to confirm their posture. The system will generate voice commands such as "straighten your back" and "relax your shoulders" in real time based on posture deviations, realizing precise interaction in the closed-eye state and greatly enhancing the immersion and effect of meditation. Attached Figure Description

[0014] Figure 1 This is a flowchart of the present invention; Figure 2 is a schematic diagram of the specific operation of Example 1. Detailed Implementation

[0015] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0016] To effectively extract human posture features and address the problem of visual feedback interfering with meditation, this invention proposes a meditation posture adjustment method based on lightweight neural networks and acoustics. The MoveNet network is used to capture skeletal key points in real time on a mobile device, and a smart cropping mechanism is used to lock the human body region. Then, a geometric feature analysis algorithm is designed to calculate posture deviations. Finally, an acoustic enhancement module is used to map the posture deviations into voice commands, which are played in real time to guide the user in adjusting their movements, without requiring visual intervention.

[0017] See appendix Figure 1 The present invention provides a meditation posture adjustment method based on lightweight neural networks and acoustics, which specifically includes the following steps: 1) Video stream input and intelligent preprocessing 1.1: Capture real-time video streams via mobile device cameras; 1.2: A smart cropping strategy is adopted. If it is the first frame of the image, the whole image is detected. If it is a subsequent frame, the region of interest (ROI) containing the human body is calculated and cropped based on the torso center point detected in the previous frame. The image size is then adjusted to the standard resolution required by the MoveNet model (such as 192x192 or 256x256) and normalized.

[0018] 2) MoveNet Feature Extraction and Keypoint Decoding 2.1: The preprocessed tensors are input into the MoveNet lightweight network, and features are extracted through the MobileNetV2 backbone network; 2.2: Using a CenterNet-based prediction head, predict human body center heatmap, keypoint regression offset, and keypoint heatmap in parallel; 2.3: Direct decoding output includes two-dimensional coordinate information of 17 key points such as nose, shoulder, elbow, wrist, hip, knee, and ankle.

[0019] Attitude geometry modeling and deviation calculation 3.1: Construct the core posture vector for meditation based on the coordinates of key points, including the spine vector (from the midpoint of the hip to the midpoint of the neck), the shoulder vector (from the left shoulder to the right shoulder), etc. 3.2: Calculate the geometric angle and positional deviation between the above vector and the standard meditation posture (such as vertical spine and horizontal shoulders).

[0020] 4) Match instructions from the speech library to the degree of deviation and provide auditory guidance. 4.1: If the angle between the spinal vector and the vertical direction exceeds the threshold, it is judged as "hunchback" or "lateral tilt"; 4.2: Retrieve the corresponding audio file (e.g., "Please straighten your back") and add it to the playback queue.

[0021] 5) Monitor and adjust dynamically until the posture meets the standard.

[0022] 5.1: Play voice commands in real time and continuously loop the above detection steps. If the user's posture returns to the normal range after adjusting according to the voice, stop the voice prompt or play a short confirmation sound (such as "Good posture") to achieve real-time interaction with eyes closed. Example

[0023] See Figure 2 This embodiment provides a meditation posture adjustment method based on lightweight neural networks and acoustics, the specific process of which is as follows: First, the acquisition device obtains a real-time video stream of the user meditating and inputs the video stream into the perception and feature layer. The preprocessing module performs size unification, image normalization, and human body region extraction on the video frames, and then inputs the processed image into the lightweight neural network model MoveNet, which outputs the two-dimensional coordinate information of key points of the human body. The key points include at least one or more of the following: head, shoulder, elbow, wrist, hip, knee, and ankle.

[0024] The logical decision layer constructs human geometric features based on the coordinates of the key points. Specifically, it calculates skeletal vectors according to the positional relationships between adjacent key points, and further calculates posture feature parameters such as the angles between limb segments, the height difference between the left and right shoulders, and the tilt angle of the torso relative to the vertical direction. These posture feature parameters are used to characterize the overall state and local deviations of the user's current meditation posture.

[0025] The knowledge base pre-stores standard meditation posture feature vectors and voice command mapping tables. The standard meditation posture feature vectors are used to characterize the relative positional relationships between key points, the range of limb angles, and the direction of the body's central axis in a standard seated meditation state; the voice command mapping table is used to establish the correspondence between "deviation location - deviation degree - prompt voice".

[0026] During the differential analysis phase, the posture feature parameters calculated in the current frame are compared with the standard meditation posture feature vectors in the knowledge base to obtain the deviation values ​​for each part. For example, when the height difference between the left and right shoulders exceeds a preset threshold, it is determined as shoulder imbalance; when the offset angle of the spinal axis relative to the vertical direction exceeds a preset range, it is determined as torso tilt; when the key points of the head are significantly offset relative to the body centerline, it is determined as head posture deviation.

[0027] To avoid misjudgments caused by momentary jitter, changes in lighting, or short-term occlusion, this embodiment further performs stability analysis on the detection results of multiple consecutive frames. Only when the same deviation occurs continuously within several consecutive frames and the deviation value reaches a set threshold is it determined as a valid erroneous posture that needs to be corrected; momentary anomalies that only occur in a single frame or a few frames are filtered out and do not trigger voice feedback.

[0028] Once a valid incorrect posture is identified, the system generates a corresponding correction instruction based on the voice instruction mapping table and sends it to the interactive feedback layer. For example, when the system detects that the user's left shoulder is higher than their right shoulder, it can generate a voice prompt such as "Please relax your left shoulder and keep your shoulders balanced"; when the system detects that the user's back is leaning forward or bent, it can generate a voice prompt such as "Please straighten your back slightly and keep your spine naturally extended". The voice prompts are processed by the acoustic enhancement module and played in real time to guide the user to adjust the corresponding parts.

[0029] After the voice feedback is issued, the system continues to collect the user's video stream and repeats the above-mentioned key point detection, geometric feature construction, differential analysis, and stability determination processes, forming a closed-loop cyclic monitoring mechanism. If the user's posture deviation is detected to decrease to within a set range, the current corrective voice prompt is stopped; in some implementations, encouraging voice messages, such as "Your current posture is good, please continue to maintain it," can also be played to enhance the user's immersion and willingness to continue adjusting.

[0030] To make the voice feedback more natural and avoid interference caused by frequent triggering, this embodiment can also perform linear smoothing or rhythm control processing on the voice commands generated in continuous time, so as to reduce the frequency of repeated broadcasts and improve auditory comfort, thereby improving the continuity of the meditation posture guidance process and the user experience.

[0031] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any equivalent substitutions, improvements, or modifications made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for adjusting meditation posture based on lightweight neural networks and acoustics, characterized in that, The method includes the following steps: Step 1: Collect real-time video stream data during the user's meditation process and perform image preprocessing to obtain a standardized input image sequence; Step 2: Use a lightweight convolutional neural network model to extract features from the input image sequence, detect and output the coordinate information of key points of the human skeleton; Step 3: Construct the geometric feature vector of the current pose based on the coordinates of the skeletal key points, and calculate the angles and relative positions of the key parts; Step 4: Compare the geometric feature vector of the current posture with the preset standard meditation posture library, and calculate the posture deviation value and deviation type; Step 5: Match the corresponding voice action command based on the posture deviation type, and generate guiding audio using the acoustic enhancement module; Step 6: Play the voice guidance instructions in real time and repeat steps 1-5 to dynamically monitor posture changes until the posture deviation value is lower than the preset threshold, thus completing the posture adjustment.

2. The meditation posture adjustment method according to claim 1, characterized in that, Step 1 specifically includes: 1-1: Obtain raw video frames of the meditation scene using an infrared or RGB camera, and use a face detection algorithm to locate the user area and crop it; 1-2: Adjust the cropped image to the 192*192 resolution required by the lightweight neural network, and perform normalization to obtain a standardized input image sequence.

3. The meditation posture adjustment method according to claim 1, characterized in that, Step 2 specifically includes: 2-1: Construct a lightweight backbone network containing depthwise separable convolutional layers to extract features from the input image sequence layer by layer, reducing the number of parameters and computational cost; 2-2: The shallow high-resolution features and deep strong semantic features are fused together, and the channel attention mechanism is used to enhance the feature response of key parts. Then, the coordinate information of the skeletal key points is obtained by directly regressing the key point heatmap using a differentiable Soft-Argmax operation.

4. The meditation posture adjustment method according to claim 3, characterized in that, Step 2-2 specifically includes: 2-2-1: Channel concatenation is performed on shallow high-resolution features and deep strong semantic features. The channel attention mechanism is used to recalibrate the weights of the feature channels to enhance the feature response to the extremities. 2-2-2: Generate a heatmap of human key points based on the enhanced features, and use a differentiable Soft-Argmax operation to directly regress the pixel-level coordinates of key points from the heatmap to obtain the coordinate information of skeletal key points.

5. The meditation posture adjustment method according to claim 1, characterized in that, Step 4 specifically includes: 4-1: Select the spine, shoulders, neck, and hips as core key points to construct the core skeletal vector of the meditation posture; 4-2: Calculate the angle between adjacent limb segments using the core skeleton vector, and analyze the posture stability in conjunction with continuous frames to obtain persistent erroneous postures, and calculate the difference between the angle between each limb segment and the corresponding angle of the standard meditation posture. 4-3: Combine the differences of multiple key parts, calculate the overall posture deviation value by weighting, and determine the body part with the largest deviation to obtain the posture deviation value and deviation type.

6. The meditation posture adjustment method according to claim 5, characterized in that, Step 4-2 specifically includes: 4-2-1: Calculate the actual included angle between each adjacent limb segment in the core skeleton vector; 4-2-2: Compare the calculated angle with the corresponding angle of the standard meditation posture to obtain the angle difference; 4-2-3: Combine continuous frame analysis to determine attitude stability and identify persistent erroneous attitudes.

Images