Unsupervised Trajectory Segmentation and Gesture Recognition through Curvature Analysis and the Levenshtein Distance

This work discusses segmentation and gesture recognition in human-driven robotic trajectories, a technique with applications in several sectors such as in robot-assisted minimally invasive surgery (RMIS) training. By decomposing entire movements-gestures-into smaller actions -sub-gestures-, we can address gesture recognition accurately. This paper extends a bottom-up approach used in surgical gesture segmentation and incorporates natural language processing (NLP) techniques to match sub-gestures with letters and treating gestures as words. We evaluated our algorithm using two different datasets with trajectories on 2D and 3D. This NLP-inspired model obtains an average F1-score of 94.25% in the segmentation tasks, an accuracy of 87.05% in the learning stage, and an overall accuracy of 88.79% in the fully automated execution. These results indicate that the method effectively identifies and interprets new surgical gestures autonomously without the need for human intervention.