Online visual gyroscope for autonomous cars

Abstract

Knowing the exact position and rotation is a crucial necessity for the navigation of autonomous robots. Even in outdoor environments GPS signals are not always accessible for estimating online rotation and position of robots. Also inertial aided navigation methods have their own defects such as the drift of gyroscope or inaccuracy of accelerometer in agile motions and environmental sensitivity of compass. In this article, we have introduced a novel online visual gyroscope that can estimate the rotation of a moving car with analyzing the images of a monocular camera installed on it. Our real time visual gyroscope utilizes an efficient method of rotation estimation between each pair of camera frames neither considering 3D points nor vanishing points. Instead, our approach assumes a fixed depth for the majority of matched key points between two frames which is more prevalent in outdoor environments like the case of autonomous car. We also analyzed different methods of extracting 2D correspondences between two frames and concluded the optimum factors for a real time implementation. Based on these determinations, we evaluated our visual gyroscope in several datasets from KITTI benchmark and showed that it can estimate the rotation with the rate of 10 frames per second and has the average drift of 0.08 degree per frame.