Monocular obstacle avoidance framework for autonomous navigation

This thesis presents a vision-based autonomous navigation framework that integrates deep learning-based monocular depth estimation with a Model Predictive Control (MPC) strategy for dynamic obstacle avoidance in indoor Unmanned Aerial Vehicles (UAVs). The proposed system addresses key challenges of operating in cluttered indoor environments where tradi- tional localization and depth sensing solutions are impractical due to hardware constraints or environmental limitations. Leveraging a fine-tuned Depth Anything V2 model, the frame- work generates dense depth maps in real time and utilizes them to construct sector-based spatial constraints within the UAV’s visual field. These constraints are incorporated into the MPC formulation to inform predictive control decisions and enable safe trajectory planning. A visual feature extraction module based on marker detection provides the reference trajec- tory for visual servoing, while the UAV continuously updates its path to avoid obstacles using dynamic depth constraints. The system was experimentally validated on a Tello quadrotor in various indoor scenarios, including static target alignment, dynamic target tracking, and ob- stacle intrusion. The results demonstrate reliable visual tracking, real-time depth estimation reaching 40 Hz via TensorRT optimization, and successful avoidance behavior under com- plex visual conditions. The contributions of this work include the design of a lightweight real-time perception-to-control pipeline, the integration of DL-based depth constraints into an MPC framework, and the demonstration of safe, closed-loop UAV navigation in dynamic environments. Although the system is designed for aerial robots, its modular architecture and sensor-driven control strategy generalize to other mobile robotic platforms. Ultimately, this framework equips mobile robots with advanced perception capabilities that are essential for achieving higher levels of autonomy in complex and unstructured environments.

https://orcid.org/0000-0003-3060-7033