Comparative Study of Traditional and Deep Learning Feature Detectors and Matchers for Land Vehicle Monocular Visual Odometry

Visual odometry (VO) enables autonomous navigation for land vehicles by estimating motion through visual data. This paper presents a comparative study of traditional feature detection and matching methods versus deep learning-based approaches in the context of land vehicle visual odometry. In feature detection, we evaluate classical techniques, namely, SIFT and ORB, alongside the state-of-the-art deep learning frameworks SuperPoint and DISK. In feature matching, we compare traditional nearest neighbor matching methods, namely, Fast Library for Approximate Nearest Neighbors (FLANN), Second Nearest Neighbor (SNN), Second Mutual Nearest Neighbor (SMNN), First-to-First Geometrically Inconsistent (FGINN) to the deep learning based Adaptive Locally-Affine Matching(ADALAM), and LightGlue. The study emphasizes performance metrics, including computational efficiency and trajectory estimation accuracy, under varying environmental conditions common in driving scenarios. Our findings highlight the strengths and limitations of each method, demonstrating that while deep learning methods excel in challenging scenarios, traditional techniques remain competitive in specific applications due to their computational simplicity. This work provides valuable insights into selecting and designing feature detection and matching pipelines for robust land vehicle VO systems. Our work provides insights that underscore the importance of tailoring detector-matcher combinations for diverse driving scenarios, balancing accuracy and computational efficiency. We made our implementation open-source at: https://github.com/olaayman/MonocularVisualOdometry.