Occupational diseases have been a significant cause of organ harm in the developed world in recent decades. The loss of motion is a major symptom and may also afect people's elbows, wrists, neck, knees, and other joints. It is preferable to treat work-related illnesses with exercise treatment. Recent advances in computer vision and machine learning have led to the suggestion of computationally more afordable alternatives. However, it is still dificult and insuficiently researched in the scientific-technical literature for a health professional to use artificial intelligence to monitor a patient while engaging in physical rehabilitation exercises. The contribution of this study aids humans by providing a visual system to analyze exercise posture and provide feedback on how to improve it. A human's stance is evaluated with an AI-based pose estimation method. It helps the user determine what kind of exercise they are doing by providing information like "Degree of Motion Required" for each move they make. To perform a posture analysis, this system uses OpenCV written in Python. Posture analysis is performed here on the live video feed. As the user exercises, this device analyzes their upper extremity posture in real-time and delivers feedback.
In 2020, 21% of all injuries and diseases inciting days from work resulted from Work-related
Musculoskeletal Disorders (WMSD), with a center of 14 days from work diferentiated and 12 days for any
excess nonfatal injury and afliction cases. In Malaysia, 61% of laborers rely on a PC-based working
environment. The side efects of outer muscle issues have multifactorial gamble factors like decision
scope, mental interest, social assistance, work flimsiness, and so forth., Musculoskeletal disorders emerge
in ofices, farming, study halls, businesspeople, and so on [
The specialists like modelers, engineers, engineers, creators, and even analysts need to invest more
energy in PCs. A few laborers must spend a similar stance over a longer term of more than 10 hrs. It is
because of high work pressure; that work environment design doesn’t meet the prerequisite for the
idea of the work; in the late days, more specialists are required to do ShiftWise design because of the
excellent eficiency. These issues make the issue in the human body parts like muscles, tendons, joints,
bones, ligaments, ligaments, and nerves [
Vision-based sensors have recently been implemented in the field of activity monitoring. They
can collect reliable skeletal data. There have also been major developments in the fields of Computer
Vision (CV) and Deep Learning (DL). Because of these causes [
Musculoskeletal disorders primarily influence the two pieces of our entire body. One is chest area
parts like the Shoulder, Neck, Wrist, Fingers, and Joints in the hand, and the other is lower body
parts like the Hip, Leg, knee joints, Ankle, and foot. In keeping an eye on the hardships of human
oversight and evaluation of proactive errands, we base on the innate issue of a human stance appraisal
that recognizes the spots of human body key joint centers (shoulder, elbow, wrists, etc.) from a lone
picture, a movement of images by a singular camera, or diferent pictures from various cameras [
Our contributions to this article are deliberated as follows: • Predicting the upper limb exercise pose without any pre-trained model. Generation of exercise pose from real-time video feed with the help of MediaPipe pose estimation library and computer vision method. • Generation of real-time feedback on the degree of angle required for the particular body part.
This ensures that the users do the exercise correctly.
An overview of the paper’s structure is provided below. In Section 2, we’ll talk about the research that’s been published on the subject of AI-powered exercise therapy systems. Section 3 will dive deeper into the proposed methodology, covering pre-processing, the pose estimate concept, angular determination from participants, and corrective feedback during exercise. The analytic results and discussion of various upper extremity rehabilitation exercises are presented in the fourth section, and our conclusion and suggestions for future research are shown in the final section.
Computer vision methods show promising answers for human posture assessment with the help of
Android handsets [
Movement boundary estimation is fundamental for grasping creature conduct, investigating the laws
of item movement, and concentrating on control techniques. These days, high-level computer vision
because of AI innovation upholds markerless articles following 2D recordings [
This framework is very simple and easy to use. The feedback for the exercises is given to the users in real-time. They can correct the pose instantly and perform the exercise perfectly. This model can be utilized in exercise centers as they have enrollment plans which the clients at that point pay. therefore, the model can be given to the participating clients.
The model gives live visuals during the whole exercise which prompts an autonomous excursion accordingly diminishing general cost. In this work, we focus on the solutions for the following important challenges in AI-based exercise pose analysis of existing systems.
• The process of gathering and annotating a lot of data is necessary to train and test AI models, but
it may be time-consuming and expensive [
Because of this, creating models that can precisely analyze a variety of exercises is challenging
[
Upper extremity exercise poses datasets were collected from more than 200 samples from the 12 various healthy subjects. We gathered the exercise pose images from the humans concentrating on the Wrist, 3D Automated Joint Assessment (3D AJA) and Kinect Software Development Kit (SDK) performance are compared in this literature OpenCV-based AI trainer model for bicep curl fitness exercise Performance comparisons of CNN, CNN-LSTM, and CNN GRU have been done in this work
Random Shoulder Forest (RF) Flexion, with Open- Shoulder source abduction, Computer and elbow vision algo- flexion rithm upper limb exercise poses Bicep curl counting for fitness exercise. cv2, Media pipe, numpy, and OpenCV is a crossplatform module which used in this work Hybrid Deep Learning used to train the model
Shoulder abduction, adduction, internal rotation, Elbow flexion, and extension exercise poses Yoga pose
Datasets Kinect and Camera OpenPose based pose estima- technique tion through open- to represent pose the human pose in 3D
The predic- These exercise poses tion accuracy for fitness applications of this frame- only not applicable for work is 91% rehabilitation training. Neck, and Elbow joints. This section describes the methodology of our proposed system. Figure 2 shows the process flow of our system to estimate the pose for the various upper extremity body exercises.
No standard procedure is followed to collect the datasets from the subjects with their own interest. Figure 3 shows the sample datasets collected from the subjects.
In the following subsections, we deliberated the various operations involved in the process of preprocessing, pose estimation concept, and determination of angles from the subjects and provided feedback.
In our work, we use the spyder (python 3.9) IDE to assess the pose. Firstly, data are called inside the process of operation with the help of the "cv2" module. Here, the "media pipe" module is used to find the landmarks in the human body. The input datasets are in the form of BGR (Blue, Green, Red), which needs to be converted into RGB using the "cv2.cvtColor" function for assessing the pose in real-time. The pixel size is fixed as 1280x720 using the function called "cv2.resize" to assess the pose angles clearly in the datasets.
The pose module is the process that helps to detect the pose by using the 33 landmarks mentioned in Fig. 1. Here, we define the class called "poseDetector" to find Pose, Position, and Angle. This class is derived with the initialization of static image mode, model complexity, smooth landmarks, enable segmentation, detection confidence, and tracking confidence from the pose module to detect the pose accurately. Finding the pose can be done by fixing the landmarks in the dataset and connecting the landmarks with the attribute "POSE_CONNECTIONS". The position of the landmarks is determined by using the "list" variable in the form of an array. In real-time, the position of the landmarks is found by using the "list".
To find the angle for the pose we used the 3 points namely, 1, 2, and 3 from the “list” of the particular landmark for the particular pose. The angle measurement of these three points is calculated by using the following eq. 1 in the Python programming language.
= ℎ.(ℎ.2 (3 − 2, 3 − 2) − (− ℎ.2 (1 − 2, 1 − 2))) (1) where, 1, 2, 3, 1, 2, 3 are the body joints points for diferent exercise poses.
Finally, the pose assessment in real-time can be achieved through the comparison of angles from the pose detector class displaying the type of exercise and providing the degree of motion required to perform the particular exercise properly in the heuristic method. We fixing the angle variation based on the reference gathered from the system execution. Reference angle values are taken by performing the execution on the dataset and measuring the angle through the display.
This work utilizes the body landmarks with the help of the “media pipe” module which is shown in Fig. 1. The analysis of the pose can be achieved by finding the angle by using eq. 1. The following is Table 2. Shows the Poses and Parameters used to analyze the exercise pose in real-time.
Figure 4 shows the sample execution of our system’s output with its corresponding exercise pose label, if the exercise is performed by the user correctly.
The main focus of this work is to provide real-time visual feedback on the correctness of exercise poses performed by the participants for the various upper limb exercise poses. The variation of angle deflection of body parts is also displayed for the clarification of angle deviation still to be done by the user. This system doesn’t require a specialized capture device like a Microsoft Kinect sensor, or RGB depth camera to analyze the correctness of exercise pose. Capturing of human action is done with the webcam on the desktop or laptop. Even though, it provides good results and guidance there are still some limitations to consider for future work.
1. Accuracy of the system needs to be analyzed to improve the quality exercise pose classification outcome of the system. 2. We plan an in-home lower body reclamation system that licenses patients to finish recuperation without assistance from any other person at home through Android mobile. 3. Occlusion of body parts during the real-time analysis is still a big challenge in the system. 4. Analysis of musculoskeletal patients’ exercise pose based on the American Academy of Orthopaedic Surgeons (AAOS) standard based on goniometer angle measurement for the prescribed exercises by the physiotherapist.
The proposed system is very simple to use and provides good results to the performer. This framework provides precise results in the output to analyze the complex exercise pose concentrating on musculoskeletal disorder patients. Utilizing this system will provide good guidance to know about the user’s exercise pose.
The authors have not employed any Generative AI tools.