Self Co-articulation Removal in Mid-air Gesticulated Trajectories via a Sequence-to-Sequence Based Classification Approach using LSTM

The inability of the acquisition device to differentiate the actual gesticulation (true stroke) from intentional movements (self co-articulation) results in the formed trajectory to incorporate both strokes. The presence of self co-articulated strokes results in confusion among the gesticulated characters at the recognition stage. Removal of these self co-articulated strokes needs to be invariant to motion characteristics (velocity/acceleration), as individuals may gesticulate at their arbitrary speed. Hence, LSTM based self co-articulation approach is proposed by considering the input gesticulated trajectory as a sequence. Motion invariant features were extracted by selecting the co-ordinates of either x-axes/y-axes based on initial gesticulation. It was observed that the proposed approach of sequence-to-sequence classification at the co-ordinate level is able to detect the self co-articulated points with a mean error rate of ∼4.03. A relative improvement of 1.82% is achieved over the state-of-the-art models used for self co-articulation removal. The proposed approach is able to overcome the dependency on hard threshold rules/motion characteristics and make the system speed invariant