Abstract:
This letter focuses on categorizing diver gestures by analyzing angle features extracted from the movements of their upper limbs without exploiting information encoded by...Show MoreMetadata
Abstract:
This letter focuses on categorizing diver gestures by analyzing angle features extracted from the movements of their upper limbs without exploiting information encoded by the hands, as is generally the case in the literature. Our approach is intended to be as generic as possible, in order to enable gesture recognition, whatever the diver's equipment, and to use the usual signs used by divers. New shallow RNN pipelines based on LSTM and GRU are proposed and evaluated with regard to a DTW-KNN deterministic baseline. For underwater gestures, a preliminary energy-based SVM separation stage is introduced to distinguish between one-arm and two-arm gestures. All classification strategies are validated using a leave-one-out protocol on a motion capture dataset comprising 14 divers performing 11 distinct gestures. The database was collected in-house with a total of 1078 individual gesture recordings. The SVM separation stage clearly improves the results, from 15% for DTW-KNN to 5% for RNNs. The best RNN leave-one-out classification accuracy is obtained for the proposed two-layer LSTM network combined with a 1D-convolution layer, and a fully connected layer, yielding a 89.5% good classification rate, compared with a 92% rate using the DTW-KNN baseline.
Published in: IEEE Robotics and Automation Letters ( Volume: 9, Issue: 11, November 2024)