Development and Implementation of Visual Odometry for Autonomous Vehicles

Abstract

Autonomous driving technology is a system that allows vehicles to navigate without human intervention. The implementation of this technology employs a variety of sensors, algorithms, and computer systems. The principles of these systems can be broadly divided into three areas: perception, decision-making, and control. This research focuses on the perception aspect. This study aims to implement the technology of Visual Odometry, which estimates the relative motion of a vehicle using camera sensors, on an embedded board installed in an autonomous vehicle. To accomplish this technology, testing with high-performance computers and compatible hardware would yield higher performance and relatively straightforward implementation. However, embedded boards require stable power efficiency and system operation, making it essential to implement them solely in lightweight programming languages like C++. Additionally, due to limited compatibility, the use of other packages is restricted, posing significant challenges in development. In conclusion, we have established a system environment that enables the implementation of Visual Odometry via an embedded board, specifically using the Nvidia Drive AGX Xavier. The operating principle involves extracting feature points from real-time camera footage. Then, by referencing previous and current continuous frames, the relationship between two images is deduced through a process of feature point matching and matrix operations. This approach allows for determining the rotation and direction of movement of autonomous vehicles. The results obtained from this process can subsequently be used to generate a vehicle's trajectory. The efficacy of this system was demonstrated through experiments involving straight and round-trip driving scenarios.