Depth Estimation from Monocular Infrared Images for Autonomous Flight of Drones

Tomoyasu Shimada, Hiroki Nishikawa, Xiangbo Kong, Hiroyuki Tomiyama

Abstract: In this work, we propose a depth estimation method from an image of an infrared camera to avoid collision for night flights of a drone. The highest flight speed of a drone is generally approximate 22.2 m/s, and long-distant depth information is crucial for night flights since if long-distance information is not available, the drone flying at high speeds is prone to collisions. However, depth cameras with long measurable distance are too heavy to equip on a drone. This work applies Pix2Pix, which is a kind of Conditional Generative Adversarial Networks (CGANs), and it generates depth images from an infrared camera. The models are trained with taking advantage of AirSim, which is one of the flight simulators. Airsim can take both infrared and depth images over a hundred meters in the virtual environment, and our model generates a depth image that provides longer distance information than the images taken by a common depth camera. We evaluate the effectiveness of our proposed method in terms of PSNR and SSIM using test images in AirSim. In addition, the proposed method is utilized for flight simulation to evaluate the effectiveness to collision avoidance.