OV2SLAM on EuRoC MAV Datasets: a Study of Corner Detector Performance

Theoretically, SLAM (Simultaneous Localization and Mapping) systems acquire information from its environment with sensors, extract landmarks from the received data and estimate its location on a map based on the sensor measurements. EuRoC datasets is a batch of visual-inertial measurements from embedded stereo camera and inertia measurement unit in a Micro Aerial Vehicle (MAV). The MAV flights include eleven itineraries and took place in indoor environments: an industrial environment and two motion capture rooms. OV
²
SLAM (Online and Versatile Visual SLAM) is an open-source visual feature points-based SLAM methods that is remarkably efficient. A feature points-based method extracts and tracks keypoints because they represent stable features. The native keypoint detection method in OV
²
SLAM tested in the EuRoC MAV datasets is a well-known KLT (Kanade-Lucas-Tomasi) corner detector. Nevertheless, no other detector was experimented on this SLAM method. This paper enables the investigation of which corner detector is optimum for OV
²
SLAM method on the EuRoC MAV datasets. Overall, the experiments are led on 10 itineraries containing 28 058 stereo-pair images in all. Thus, by varying the parameter of the Gaussian influencing the detection of the keypoints, a global score based on different statistics is calculated in relation to the ground truth to classify which pair detector/parameter is optimal on these datasets.