Depth, Spatial, and Temporal Features for Visual Odometry in Unstructured Agricultural Environments

Abstract

Unstructured agricultural environments
comprises several elements that present significant
challenges to visual odometry (VO) methods. Deep
learning-based VO methods have been proposed in
state-of-the-art which have been demonstrated out-
standing performance in structured environments even
superior to conventional methods, nevertheless, these
methods fail when face to unstructured environments
such as the agricultural. In this work, we propose
deep learning based VO model that exploits depth,
spatial, and temporal features. To do this, the proposed
model comprises two image processing pathways: the
scene and depth pathway. The features extracted in
these pathways are then fused to computes the relative
pose given as R and t components. We conduct
experiments evaluating the proposed model employing
the agricultural Rosario dataset. Results show that
the use of depth features improves significantly the
performance of proposed model, obtaining consistent
and coherent estimated trajectories in training and
testing sequences.