Autonomous Drone Racing Via Vision-Only Monocular SLAM and Submapping

Abstract

We present an Unmanned Aerial System (UAS) for Autonomous Drone Racing (ADR) in which state estimation relies exclusively on monocular Simultaneous localization and Mapping (SLAM). In contrast to approaches that depend heavily on known gate geometry and tightly coupled visual–inertial fusion, we propose a lightweight vision-only solution in which gate information is used solely during map initialisation to recover the metric scale. To mitigate long-term drift, the proposed system employs a submap decomposition strategy that exploits prior knowledge of the race-track layout. For flight control, a cascade PID architecture was implemented to generate roll, pitch, and yaw angle-rate commands transmitted to the flight controller. The complete system operated fully onboard the quadrotor without GPU acceleration at a frequency of 20 Hz,
achieving speeds of up to 3 m/s whilst successfully completing the race track autonomously.