Monocular Depth Estimation And Collision Prediction For Quadrotors

This thesis introduces a novel approach for quadrotor navigation using monocular depth estimation integrated with model predictive control. The method involves embedding the estimated depth as a latent variable in the quadrotor's state and learning the dynamics of this latent state to predict collision probabilities. Key challenges addressed include developing a real-time, low-power framework suitable for battery-constrained aerial vehicles, using only lightweight cameras and an inertial measurement unit. The system is designed to be resilient against factors like motion blur and quadrotor dynamics such as aerodynamic effects and motor vibrations. Experimental results demonstrate the system's effectiveness in precise navigation and obstacle avoidance.