On feature selection in automatic detection of fitness exercises using LSTM models

As constantly stated by the World Health Organization, physical activity is extremely important for a healthy aging, but how exercises are made is as important as how much activity is made. A large variety of wearable devices capable of sensing people movement is available on the market. Automatic detection and classification of fitness activity is also possible, leveraging artificial intelligence (AI) algorithms. In this paper, some ideas on the impact of specific input features on AI model performance for fitness exercise recognition is reported and discussed. Then, a general classification of input features is proposed. Using a pre-recorded dataset composed of 9 exercise repetition sets performed by 7 volunteers, a LSTM network have been trained and validated using the Leave One Out Cross Validation approach. Finally, the same network has been retrained several times, varying the input parameters. Differences in classification results due to such parameters have been evaluated through the precision, recall and accuracy metrics. In particular, the precision is between 97.8% and 63.8%, whereas recall is between 98.5% and 42.3%, in line with results in literature.

Automated summary

This paper discusses the impact of specific input features on the performance of an artificial intelligence model for fitness exercise recognition, and proposes a general classification of input features. A LSTM network is trained and validated using a pre-recorded dataset, and the effects of varying input parameters on classification results are evaluated.