In indoor navigation and visual odometry, cameras may be used as an aiding algorithm for inertial navigation. Fast and accurate image matching is an important task used in various applications in computer vision and visual odometry. In this research, the performances of all recently available detection, description, and matching techniques were compared. The detection techniques included HARRIS, GFTT, SIFT, SURF, STAR, FAST, ORB, MSER, Dense, and SimpleBlob. Feature description techniques included SIFT, SURF, ORB, and BRIEF. Finally, image matching techniques included BruteForce, BruteForce-L1, BruteForce-Hamming, BruteForce-HammingLUT, and FlannBased. The comparison was made between different kinds of geometric distortions and deformations, such as scaling, rotation, fish-eye distortion, and shearing. In this research, the lighting conditions and the shadowing effects were not taken into consideration.
To perform this task, different types of transformations were manually applied to original images and computed with the matching evaluation parameters, such as the number of detected keypoints in images, the processing time, and the matching accuracies for each algorithm to show which algorithm was the best and most robust against these distortions. This work provided us with a perspective on and contributed to the field of image-based indoor navigation systems to recommend further research. The results showed that the ORB detector, the ORB descriptor, and either the BruteForce-Hamming or the BruteForce-HammingLUT matchers were favored to be used in indoor environments.
