Visual odometry is a widely-used approach for the relative localization problem in autonomous vehicles. However, achieving high-accuracy pose estimation in a dynamic driving environment while meeting safety requirements is challenging. Existing geometry-based visual odometry algorithms require individual design and fine-tuning. Therefore, we propose a new monocular visual odometry system based on a neural network for feature extraction and matching. The efficiency of measurement is a primary consideration due to safety requirements in complex environments, and reducing computational cost is crucial for enabling real-time processing on embedded platforms with limited hardware resources. Our system estimates poses directly from video sequences, and our main contribution is an efficient neural network model that reduces computational costs while maintaining accuracy. Our proposed model is suitable for applications such as robotics and autonomous driving, which require a lightweight, low-latency network model. Experimental results show that our system is relatively accurate to ground truth, and our lightweight neural network model is a significant contribution to visual odometry by reducing computational costs.
Lightweight Deep Learning Algorithm for Visual Odometry
2023-08-28
998599 byte
Conference paper
Electronic Resource
English
IEEE | 2022
|IEEE | 2004
|British Library Conference Proceedings | 2004
|Visual Odometry for Autonomous Deep-Space Navigation
NTRS | 2016
|