SmartCoach personal gym trainer: An Adaptive Modified Backpropagation approach

摘要

The enhanced performance of modern Artificial Intelligence algorithms has opened up limitless possibilities in the development of smart systems and devices. One of the most complex tasks for interactive devices is the analysis of human motion. However, using neural networks, movement can be classified and even understood. This paper proposes the use of an enhanced backpropagtion algorithm for developing a device that functions as a personal exercise trainer. It offers a full workout program, monitors the user's workout, counts down the reps and alerts the user when the move is performed incorrectly. Due to the real-time nature of the workout, an efficient algorithm must be used. The most popular one, the standard backpropagation (SBP) algorithm, updates weights using a fixed learning rate, which is optimized using trial and error. An improved algorithm, known as the Adaptive Modified Backpropagation (AMBP) algorithm, speeds up convergence by adapting the learning rate at each layer and at every epoch. In addition, the final error reached is significantly diminished by taking into consideration a new, linear error along with the standard nonlinear error. The feasibility of using AMBP to monitor and classify human motion is demonstrated by comparing system performance, specifically convergence speed and final error, with those obtained by standard algorithms.