Exercise Recognition and Repetition Counting for Automatic Workout Documentation Using Computer Vision

Francois Volschenk, Hima Vadapalli, Dustin van der Haar

Abstract: This paper aims to study various approaches using deep learning methods to perform human action recognition (HAR). More specifically, a subset of HAR focused on recognising exercises and counting repetitions using deep learning. The paper discusses two approaches used in an attempt to produce a machine-learning model that is capable of identifying certain exercises from video input. This model is then incorporated into a system that can document a person’s workout by identifying the exercises being done and counting the repetitions of each exercise. The study uses artificial training data in 3D animated videos of avatars performing the exercises. The dataset used is InfiniteRep from InfinityAI. Feature extraction and repetition counting are performed using the Mediapipe pose estimation model. An LSTM-based model and a 1D time-distributed CNN are used for exercise recognition. The models were compared on classification metrics: accuracy, precision, and recall. The LSTM-based model produced a 96% accuracy on the dataset, whereas the CNN-based model produced 97.3% accuracy on the same dataset. The CNN-based model is also capable of performing in near real-time.

External IDs:dblp:conf/hci/VolschenkVH24