Evaluation and Optimization of Feature Detectors Towards Off-Road Visual Odometry

Abstract

Feature detectors are used in many computer vision applications, including Visual Odometry (VO). However, their utilization in off-road VO remains a topic of great interest. In this body of work, software tools are developed in order to evaluate and parameter-optimize feature detectors towards real-time off-road VO. Using the tools developed, various classical as well as state-of-the-art machine learning-based feature detectors are evaluated and optimized, including their usage in a pre-existing VO implementation and analyzing the output trajectory. The analysis and results presented show that although the quality of feature detectors have an impact, optimizing them alone cannot overcome the inherent drawbacks of monocular VO approaches. Based on the analysis, recommendations of potential feature detectors that can support real-time off-road VO are made and optimized parameters of selected feature detectors are provided. Key areas of improvement for future research in the field are also identified.