Comparison of local feature extraction paradigms applied to visual SLAM

The detection and description of locally salient regions is one of the most widely used low-level processes in modern computer vision systems. The general approach relies on the detection of stable and invariant image features that can be uniquely characterized using compact descriptors. Many detection and description algorithms have been proposed, most of them derived using different assumptions or problem models. This work presents a comparison of different approaches towards the feature extraction problem, namely: (1) standard computer vision techniques; (2) automatic synthesis techniques based on genetic programming (GP); and (3) a new local descriptor based on composite correlation filtering, proposed for the first time in this paper. The considered methods are evaluated on a difficult real-world problem, vision-based simultaneous localization and mapping (SLAM). Using three experimental scenarios, results indicate that the GP-based methods and the correlation filtering techniques outperform widely used computer vision algorithms such as the Harris and Shi-Tomasi detectors and the Speeded Up Robust Features descriptor.