Ground view-based SLAM using a commercial web camera in unstructured indoor environments

This paper presents a ground view-based Simultaneous Localization and Mapping (SLAM) using a commercial web camera to construct a feature map of an unstructured indoor environment. A camera attached to a robot looks downward to exploit several advantages; being insensitive to lighting conditions; limiting visible region to be undisturbed by moving objects; extracting vertical edge features easily. In order to make full use of characteristics of indoor environment, we choose to extract a salient vertical edge feature. Feature parameters including uncertainty are estimated by a random sample consensus technique using the fact that all vertical edge features intersect at a fixed vanishing point. Through successive tracking of extracted feature, its initial location is determined by a non-linear least squares method. Experiment in a small-scale but highly unstructured environment shows a good SLAM performance under Extended Kalman Filtering algorithm that runs in real-time.