Performance Analysis of Keypoints Detection and Description Algorithms for Stereo Vision Based Odometry

This paper presents a comprehensive evaluation of keypoint detection and description algorithms for stereo vision-based odometry in dynamic environments. Five widely used methods—FAST, GFTT, ORB, BRISK, and KAZE—were analyzed in terms of detection accuracy, robustness to image distortions, computational efficiency, and suitability for embedded systems. Using the KITTI dataset, the study assessed the influence of image resolution, noise, blur, and contrast variations on keypoint performance. The matching quality between stereo image pairs and across consecutive frames was also examined, with particular attention to drift—cumulative trajectory error—during motion estimation. The results show that while FAST and ORB detect the highest number of keypoints, GFTT offers the best balance between matching quality and processing time. KAZE provides high robustness but at the cost of computational load. The findings highlight the trade-offs between speed, accuracy, and resilience to environmental changes, offering practical guidance for selecting keypoint algorithms in real-time stereo visual odometry systems. The study concludes that GFTT is the most suitable method for trajectory estimation in dynamic, real-world conditions.