Advancements and Performance Analysis of Feature Detection and Matching Methods in Visual Odometry

This study presents a comprehensive performance analysis of feature detectors and descriptors in visual odometry (VO) based on four key metrics: absolute trajectory error (ATE), relative pose error (RPE), cumulative drift, and frames per second (FPS). Five algorithm-based methods (SIFT, ROOTSIFT, ORB, KAZE, and AKAZE) and five AI-based methods (SuperPoint, DISK, ALIKED, D2NET, and Xfeat) are evaluated using two matching strategies: Brute-Force and LIGHTGLUE. Performance assessment is conducted on the KITTI dataset, demonstrating that KAZE achieves the best results in ATE and cumulative drift, while ALIKED outperforms in RPE and FPS. These findings offer valuable insights for selecting the optimal feature detection and description techniques in real-world VO applications.