PHYSICAL ACTIVITY TECHNIQUE MONITORING (PATMo) BASED POSE ESTIMATION USING CNN

Exercising is a physical activity to increase the quality of life. However, exercising may come with various injuries ranging from minor to an injury that can cause fatality. Exercising requires gathering information to perform a specific physical activity with extreme caution and safety. Immediate feedback on human pose on performing a physical activity is a prime of importance to avoid harm to the human body. This study emphasizes pose estimation using a Convolution Neural Network model to identify a person’s key points. A prototype called Physical Activity Technique Monitoring (PATMo) embedded with Mobilenet-YOLOv3 and Simple Pose Tesnet18 v1b models is developed. PATMo focuses on a single movement and angle for receiving feedback for the physical activity. PATMo utilizes two optimizers and several batch sizes for parameter tuning. The batch size 32 with Adaptive Moment Estimation optimizer has the highest accuracy of 82.66% to Stochastic Gradient Descent, but the computational time took about 12 hours. More evaluations are expected with more powerful computer and Convolution Neural Network models variants. It is a starting point for further investigation to improve the feedback time during physical.