Grid-based visual SLAM in complex environments

Abstract

This paper presents a novel approach to the real-time SLAM problem that works in unstructured indoor environment with a single forward viewing camera. Most existing visual SLAM extract features from the environment, associate them in different images and produce a feature map as a result. However, we estimate the distances between the robot and the obstacles by applying a visual sonar ranging technique to the image and then associate this range data through the Iterative Closest Point (ICP) algorithm and finally produce a grid map. Moreover, we construct a pseudo-dense scan (PDS) which is essentially a temporal accumulation of data, emulating a dense omni-directional sensing of the visual sonar readings based on odometry readings in order to overcome the sparseness of the visual sonar and then associate this scan with the previous one. Moreover, we further correct the slight trajectory error incurred in the PDS construction step to obtain a much more refined map using Sequential Quadratic Programming (SQP) which is a well-known optimization scheme. Experimental results show that our method can obtain an accurate grid map using a single camera alone without the need for more expensive.