A Comparison of Transfer Learning Inspired CNN Architectures for 2D Image Workout Recognition
Aman Rehman, Shailender Kumar, Lakshya Verma · 2024
Physical exercise is a key part of a healthy lifestyle, encompassing a wide array of activities from dance and aerobics, to weightlifting and sports. Proper form and posture are critical to ensure the safety of these exercises as well as making them effective. Accurate assessment of workout poses can provide invaluable feedback to individuals and fitness professionals alike. Performing exercises correctly also reduces the risk of injury, especially for people new to exercises. Additionally, fitness professionals can utilize this to remotely guide clients in virtual training sessions. Such systems can also act as a guide for people working out without any assistance. Deep learning approaches have been extensively used in computer vision as they show exceptional performance across a range of tasks. This may be used to automate the process of evaluating body postures during workouts. By leveraging these architectures, the aim is to develop an efficient and accurate system capable of precise pose identification. For this, we have taken a2D image dataset of 20 different workout postures. Weexperiment using 9 CNN architectures. EfficientNetB0 was found to have the greatest test accuracy of 91%.