A Comparison Between Multilayer Perceptrons and Kolmogorov-Arnold Networks for Multi-Task Classification in Sitting Posture Recognition

Prolonged incorrect sitting postures can lead to various health issues, including musculoskeletal disorders and reduced quality of life. Efficient and accurate methods for monitoring and assessing sitting postures are crucial for promoting overall well-being in our increasingly sedentary society. This study compares the performance of Kolmogorov-Arnold Networks (KANs) and Multilayer Perceptrons (MLPs) in sitting posture recognition, utilizing a multi-task classification approach to simultaneously identify upper and lower body positions. A comprehensive dataset of sitting postures was developed using MediaPipe for skeletal extraction, labeled for both body regions, and encompassing a range of common scenarios. The models were implemented and evaluated using Leave-One-Subject-Out (LOSO) cross-validation to assess generalization capability across different individuals. Results indicated that the KAN model achieved higher accuracy (97.03% for upper body and 92.11% for lower body) with fewer parameters (4,320) compared to the MLP (93.87% and 91.47% respectively, with 7,662 parameters). Although the MLP demonstrated faster inference, both models achieved accuracies exceeding 91% in LOSO validation, suggesting their potential for real-time applications. This study significantly contributes to the field of automatic posture recognition, offering insights into the applicability of KANs and MediaPipe-based pose estimation in ergonomic recognition systems and demonstrating the effectiveness of a multi-task approach for comprehensive sitting posture analysis.