Estimação de profundidade monocular baseada em redes neurais profundas para mapeamento, localização e navegação de veículos autônomos

Abstract

We present a grid map occupancy builder system for the IARA autonomous car, using depth from monocular images obtained through neural networks. Depth estimation from monocular images is still a challenge for deep neural networks, however, recent transformer-based architectures have shown promising performance, making their use in real-world robotic applications possible. In this work, we employed the transformer-based architecture GLPDepth to build a system that calculates depth from monocular images captured by an autonomous car and, with adequate pre-processing, employs this depth information to construct appropriate grid maps for autonomous vehicles. To evaluate the performance of the proposed system, we integrated it into the perception and navigation systems of an autonomous vehicle and tested it in real-world scenarios. Our results showed that monocular cameras can be used as the main sensor for autonomous operation of vehicles at low speeds. However, it is important to note that the system's performance is still limited compared to the use of LIDAR sensors in high-speed or complex and challenging environments. The proposed system shows promising performance for building occupancy grid maps for autonomous cars using monocular cameras with transformer-based deep neural networks, and can be a more economical and simpler alternative to LIDAR sensors in certain situations. However, the use of other sensors, in addition to monocular cameras, may be necessary to ensure safety in all operating condition.