Project DetailsThis project focuses on spatial segmentation of indoor environments for autonomous drone navigation, based on depth and grayscale data from a pico flexx2 camera. The goal of the project is to partition the scene into semantic classes: general geometry (floor, walls, stairs, and ceiling), obstacles, and an “unknown” pixel class added during manual annotation for regions that are difficult to classify. The work builds upon a prior project that utilized the ESANet architecture, applying a transfer learning approach while adapting the development environment. For training purposes, a new dataset was constructed, comprising approximately 450 manually annotated images captured across the Technion campus. The proposed model achieved an mIoU of approximately 69%, demonstrating good capability in identifying indoor geometry and large structures, such as stairs and wall-mounted elements. However, obstacle classification proved challenging due to the high variability of objects in this class. Future work will focus on improving accuracy through dataset expansion and further optimization, as well as adapting the method for real-time operation for autonomous indoor navigation.
