A Modified Visual Simultaneous Localisation and Mapping (V-SLAM) Technique for Road Scene Modelling

Abstract

Visual Simultaneous Localization and Mapping (V-SLAM) which involves the use of cameras to map an environment and estimate agents’ pose within that environment has become widely popular in the field of autonomous vehicles. Numerous V-SLAM schemes have been implemented which utilize various feature extraction methods, one of which is the use of Convolutional Neural Networks (CNN). One main shortcoming of existing approaches is that they do not focus on object detection of road sceneries which are characterized by their varying complexity, thus making them unsuitable for real time implementation. Therefore, this study presents a modified V-SLAM scheme for road scene modelling. The technique utilizes YOLOv4 for object detection, and uses the ORB features obtained from the objects to update the features in the main VSLAM algorithm. The results showed that the modified VSLAM technique was capable of estimating the agent’s position and orientation, map the environment. The technique gave a Root Mean Square Error of 0.11621 and a point-to-point distance of 1.1726m.