Dynamic object-aware monocular visual odometry with local and global information aggregation

Abstract

In this paper, we present a deep learning-based approach to monocular visual odometry. We propose a LCGR(Local Convolution and Global RNN) module which utilizes several independent 3D convolution layers to filter noise from features extracted by FlowNet, as well as to model local information, and a Bi-ConvLSTM layer to model time series and capture global information. In addition, our network jointly predicts optical flow as an auxiliary task by measuring photometric consistency in a self-supervised way to help the encoder for better motion feature extraction. In order to alleviate the effects of non-Lambertian surfaces and dynamical objects in the scene, a confidence mask layer is estimated and epipolar constraint is added to the training process. Experiment results indicate competitive performance of the proposed framework to the state-of-art methods.

Reference

Yiming Wan, Wei Gao*, Sheng Han, Yihong Wu. Dynamic object-aware monocular visual odometry with local and global information aggregation. Abu Dhabi National Exhibition Center (ADNEC), Abu Dhabi, United Arab Emirates (UAE), 10.25-10.28, 603-607, ICIP 2020. DOI