Unsupervised activity recognition for autonomous water drones

We propose an automatic system aimed at discovering relevant activities for aquatic drones employed in water monitoring applications. The methodology exploits unsupervised time series segmentation to pursue two main goals: i) to support on-line decision making of drones and operators, ii) to support off-line analysis of large datasets collected by drones. The main novelty of our approach consists of its unsupervised nature, which enables to analyze unlabeled data. We investigate different variants of the proposed approach and validate them using an annotated dataset having labels for activity "upstream/downstream navigation". Obtained results are encouraging in terms of clustering purity and silhouette which reach values greater than 0.94 and 0.20, respectively, in the best models.