A Deployable Privacy-Preserving Thermal-Based Obstacle Detection System for Indoor Navigation

Indoor navigation and wayfinding have a wide range of applications, from enabling robots to efficiently navigate ware-houses to enhancing mobility for people with visual impairments (PVI). These systems rely on a combination of localization, obstacle detection, and path-finding technologies. While real-time obstacle/object detection is essential for safe indoor navigation, current solutions are impractical, face privacy concerns, and have high barriers to adoption; particularly for PVI. In this paper, we introduce DareDevil, a highly deployable indoor navigation system that ensures safe navigation while preserving user privacy. DareDevil offers three main components: a comprehensive navigation dataset, a real-time thermal-based obstacle detection model, and a thermal colorization module that feeds thermal data to the model during inference. The thermal-based dataset supports seamless navigation in both indoor and outdoor environments. Our system provides trained models of varying sizes; making it compatible with a wide range of high-end IoT devices and eliminating the need for bulky or expensive processing units. We evaluate DareDevil's performance across different model sizes and thermal image colorization schemes. Our results show that DareDevil's models can achieve higher accuracy compared to the state-of-the-art real-time object detection model. Furthermore, the findings highlight the often-overlooked impact of thermal image colorization on thermal-based models.