Feature extractors evaluation based V-SLAM for autonomous vehicles

Abstract

Visual Simultaneous Localization and Mapping known as V-SLAM, is an essential task for autonomous vehicles. It can be carried out using several sensors, in particular with on board cameras. To locate a vehicle, SLAM algorithms are based on two main tasks. The first task (front-end kernel) is intended to process images in order to provide features (called also landmarks or primitives) of the perceived environment. The second task (back-end kernel) is intended for localization and environment reconstruction. This work focuses on the front-end task which uses extractors (detectors and descriptors) in a stereo-vision system. Several feature detectors and descriptors exist in the state of the art. The aim of this paper is to evaluate the possible combinations of detectors and descriptors to achieve a precise localization while considering the processing times. The study is extended to bio-inspired extractors. The evaluation is achieved with SLAM algorithms over well-known indoor and outdoor datasets. Experimental results highlight the performance of bio-inspired extractors and their potential integration in designing vision systems for real-time SLAM applications.