Seamless Indoor Navigation: Leveraging AI Computer Vision and GPS Spoofing for UGV Autonomy

Abstract

To bypass the high computational overhead and environment-specific mapping dependencies of traditional indoor localization, this work introduces a cost-effective navigation framework that enables standard autopilot controllers to operate indoors via dynamic GPS retransmission. By integrating OAK-D AI cameras for wide-area target detection with Software-Defined Radio (SDR) technology, the system generates real-time, localized GPS signals to provide seamless position inputs to commercial off-the-shelf autopilots. Experimental results demonstrate target detection with a 69% confidence floor at an operational distance of 8.5 m. Under static conditions, Kalman filtering refined the retransmitted GPS tracking accuracy from 21 cm to 13 cm within 0.6 s. Dynamic tracking trials along a complex figure-eight trajectory demonstrated that a 100 Hz non-linear EKF—fusing 5 Hz retransmitted GPS with raw IMU variable speed—effectively neutralized indoor multipath interference. This framework achieved an exceptional 2D position RMSE of just 4.28 cm, compared to a substantial 43.20 cm error yielded by a constant-speed baseline tracking architecture. This study demonstrates that dynamic GPS retransmission provides a robust, infrastructure-light alternative to complex spatial mapping, allowing standard autonomous vehicles to navigate seamlessly within GNSS-denied environments.