Automated OWAS in forestry using deep learning for posture classification

Forestry work involves significant ergonomic risks leading to musculoskeletal disorders (MSDs). While the Ovako Working Posture Analysing System (OWAS) is a standard tool for assessment, its manual application is time-consuming, creating a need for a more efficient, automated alternative. This study aimed to develop and evaluate an automated OWAS-based postural classification system using deep learning, specifically investigating whether augmenting conventional images with computer-generated connected body keypoints improves classification accuracy. A dataset of 23,000 images from forest operations was collected, from which a balanced subset of 1260 images representing 252 OWAS classes was used to create three distinct datasets: PictureOnly (PO), PictureWithSkeleton (PS), and SkeletonOnly (SO). Using transfer learning, a pre-trained ResNet-50 model was adapted for each dataset. The models were compiled with the Adam optimizer and categorical cross-entropy loss, then trained for 50 epochs using a learning rate of 0.0001 and a batch size of 32. Performance was evaluated using accuracy, precision, recall, and F1-score on a small, perfectly balanced test set of 252 images (one per class). The ResNet-50 model trained on the PS dataset outperformed the others, achieving an accuracy and F1-score of 99.8%. The model based on PO data achieved an accuracy of 97.8% and an F1-score of 97.3%, while the model developed with SO data showed a significant drop in performance with 66.2% accuracy and a 64.3% F1-score. The results indicate that combining original images with computer-generated connected body keypoints can increase accuracy by 2-2.5%, though statistical significance was not explicitly tested given the discrete nature of the subset. However, excluding the image's background context significantly reduces classification performance. This automated classification method can enhance postural assessments and improve worker safety and health in forestry and related fields. • Body keypoints and connectors were overlapped on conventional images. • This strategy improved classification performance by state-of-art transfer learning. • ResNet-50 provided an exceptional 99.8% in accuracy and F1-score. • The model performed exceptionally on unseen data.