Activity Recognition from Skeleton and Acceleration Data Using CNN and GCN
Most of the existing methods of activity recognition are based on single-label classification, however, these methods cannot be used in this challenge which focuses on multi-label classification-based micro-activity recognition. To address this, we propose a GCN model using the binary cross-entropy loss function, which enables multi-label classification and achieves an average accuracy of 83.1% on the Cooking Activity Dataset. In addition, to utilize the advantages of multi-modal data, we propose a joint training CNN model that combines the acceleration and skeleton data together. Finally, the proposed CNN model achieves an average accuracy of 82.8% for macro-activity recognition on Cooking Activity Dataset.
