Deep adaptation network for cross-domain sitting posture recognition using pressure sensor arrays

Accurate and timely detection of improper sitting postures is essential for mitigating potential health risks. Numerous studies have explored the use of pressure sensing to monitor sitting postures, but few have systematically evaluated intra-subject and inter-subject variations in designing a practical sitting posture recognizer. To address this gap, we propose a cross-domain sitting posture recognition framework based on the transfer learning paradigm, aimed at reducing the distribution discrepancy between source-domain and target-domain pressure sensor data. The proposed framework incorporates an adaptation neural network to automatically extract discriminant features and minimize both marginal and conditional distribution discrepancies between domains in the learned latent feature space. To assess the effectiveness of the framework, comparative experiments are conducted against five traditional transfer learning algorithms and one deep transfer learning model. Performance metrics, including accuracy and F1 score, are evaluated on the collected experimental data under two typical cross-domain scenarios (i.e., cross-subject and cross-time). The results indicate that cross-domain sitting posture recognition is significantly impacted by degraded accuracy and that transfer learning helps mitigate the issue. Among the evaluated methods, the end-to-end deep transfer learning techniques generally outperform traditional transfer learning techniques, demonstrating superior cross-domain performance.