LatinX in AI at Computer Vision and Pattern Recognition Conference 2022

Abstract

Autonomous navigation is a challenging task that requires solving individual problems such as the camera pose. Even more, if agile motion is performed in the navigation, the problem becomes complex due to the necessity to know the pose (trajectory generated) as fast as possible to continue with the agile motion. Several works have proposed approaches based on geometric algorithms and deep learning-based solutions that reduce error in estimation. Still, the prediction is performed at 30Hz on average in most cases. It is desirable to increase the frequency of operation to allow cameras with higher frame rates to capture the motion correctly. Motivated by the latter, we propose a two-stream network that allows predicting pose at a frame rate up to 78f ps with an average relative error of 2.21m.