NEURAL NETWORK VISUAL ODOMETRY BASED FRAMEWORK FOR UAV LOCALIZATION IN GPS DENIED ENVIRONMENT

Abstract

This paper presents a vision-based localization framework based on visual odometry.Visual odometry is a classic approach to incrementally estimate robot motion even in GPS denied environment, by tracking features in successive images. As it is subject to drift, this paper proposes to call a convolutional neural netwok and visual memory to improve process accuracy. In fact, our framework is made of two main steps. First, the robot builds its visual memory by annotating places with their ground truth positions. Dedicated data structures are made to store referenced images and their positions. Then, during navigation step, we use loop closure corrected visual odometry. A siamese convolutional neural network allows us to detect already visited positions. It takes as input current image and an already stored one. If the place is recognized, the drift is then quantified using the stored position. Drift correction is conducted by an original two levels correction process. The first level is directly applied to the estimation by substracting the error. The second level is applied to the graph itself using iterative closest point method, to match the estimated trajectory graph to the ground truth one. Experiments showed that the proposed localization method has a centimetric accuracy.