DeepPilot: A CNN for Autonomous Drone Racing.

Area of Science:

• Robotics
• Artificial Intelligence
• Computer Vision

Background:

• Autonomous Drone Racing (ADR) requires sophisticated navigation systems.
• Existing ADR solutions often rely on a pipeline of gate detection, drone localization, and flight control.
• Deep Learning (DL) has shown promise in gate detection and localization, but a greater role for DL is desired.

Purpose of the Study:

• To propose DeepPilot, a novel Convolutional Neural Network (CNN) approach for autonomous drone navigation in racing environments.
• To enable a drone to predict and execute flight commands directly from camera input for autonomous racing.
• To investigate the effectiveness of a purely DL-based system for the ADR challenge.

Main Methods:

• Developed DeepPilot, a CNN that takes camera images as input and outputs flight commands (roll, pitch, yaw rates, and vertical speed).
• Integrated DeepPilot's output with the drone's inner controller for autonomous navigation through racetrack gates.
• Evaluated DeepPilot in simulated racetrack environments, including a temporal approach using mosaic images of consecutive frames.

Main Results:

• DeepPilot successfully navigated simulated racetracks multiple times, demonstrating repeatability.
• The system achieved an average processing speed of 25 frames per second (fps).
• The temporal approach, using mosaic images, enhanced motion trend learning and flight command prediction.

Conclusions:

• A purely Deep Learning-based artificial pilot, DeepPilot, is feasible for Autonomous Drone Racing.
• DeepPilot's ability to predict flight commands directly from visual input simplifies the ADR pipeline.
• The temporal approach enhances DL-based drone navigation by providing a form of local memory.