Evaluation of IGFTT Keypoints Detector in Indoor Visual SLAM

Abstract

Features detectors are key parts of a Visual System of Simultaneous Localization and Mapping (VSLAM). Matching Points of interest (POI) in the image are used by visual odometry to estimate the movement of a robot through camera movement, and to recognize places already visited. It is extremely important to VSLAM the quality evaluation of these detectors, in order to determine the best one to be used in a given context, because the quality of the map generated by the system depends on the POI. Recently introduced feature detector known as Improved Good Features to Track (IGFTT) presents promising results. This work evaluates the IGFTT detector in the scope of a Visual SLAM system in internal environments, comparing it with popular detectors in the academic community, such as SIFT, SURF, FAST, AGAST, GFTT. A public database is used as a benchmark with an open source VSLAM system. The quality of each detector is determined using as metric by the VSLAM pose and trajectory errors, the runtime (detection, description and POI correspondence), the repeatability rate of the detected POI, the probability of their survival in future frames and the number of successive nodes in the scene graph. The tests identify the good qualities and deficiencies of each analyzed detector. In the overall ranking, the SURF comes first, secondly the IGFTT2, a proposed refinement of the IGFTT, and thirdly the original IGFTT.