Intersection waypoint identification for vision based indoor navigation of nano-scale drones

Abstract

Nano-scale drones represent a promising platform for indoor inspection and search operations in GPS-denied, constrained environments, as their compact form factor enables access to narrow spaces. However, they require lightweight, resource-efficient autonomous navigation algorithms. Vision-based navigation offers an infrastructurefree substitute to GPS and external positioning systems. Simple vision-based navigational algorithms such as ring attractor, optical flow-based approaches, suffer from drift and cumulative error over time. Introducing semantic waypoints provides a mechanism to mitigate such drift. This work proposes an egocentric, vision-based waypoint navigation framework that formulates waypoint detection as an intersection understanding problem, cast as a concept learning task. We first formulate the intersection identification problem as a multi-class classification. To ensure a model can be trained with limited data and still preserve the conceptual understanding, each intersection is decomposed into its fundamental directional components: left, right, and forward. This novel multi-label classification is advantageous because it can extend to zero-shot inference and complex junction types. Further, we augment the limited real data with synthetic visual data collected from a safer digital twin environment which replicates the real-world scenes. We design five experimental setups in which model weights are initialized from Kaiming He, ImageNet or synthetic pre-trained weights obtained using the Digital Twin dataset. We show that, (i) Intersections can be used as waypoints for the navigation of nano-drones. (ii) Multi-label classification outperforms multi-class formulation in intersection identification. (iii) Model trained using digital twin data alone can yield an F1 score of 0.6 in real-world predictions, compared to 0.5 from a random classifier. (iv) Introducing limited real-world training data can bring the model up to near-perfect accuracy with faster convergence. (v)The multi-label classification of junctions provides insightful visual explanations ensuring reliability and generalizability of the conceptual framework. Finally, the proposed framework is effective in environments such as libraries and supermarkets where structured aisle naturally exists and GPS is unavailable. This method offers an infrastructure and drift-resistant solution for the navigation of nano-scale drones by leveraging intersections as semantic waypoints.