Vision-based Obstacle Avoidance Based on Monocular SLAM and Image Segmentation for UAVs

Abstract

The capability of a unmanned aerial vehicle (UAV) to identify an obstacle in its path and go around (avoid) it is a critical requirement that must be met before introducing UAVs into the national airspace, especially in areas close to buildings and people. Furthermore, it is highly desirable that the obstacle avoidance capability be independent of GPS which could be unavailable or degraded. Therefore, developing reliable obstacle avoidance techniques in conjunction with solutions for autonomous navigation in the absence of or presence of degraded GPS is critical for operating UAVs in urban environments. The navigation problem can be solved in the framework of the Simultaneous Localization and Mapping (SLAM) problem. In the absence of GPS, the ability to solve the SLAM problem in real time using vision as the primary sensing mechanism onboard the UAV has been demonstrated. SLAM solutions involving just a single camera, referred to as monocular SLAM, have recently been proposed in literature. Monocular SLAM solutions are attractive in context of small UAVs with regard to meeting weight, cost and power constraints. In general, monocular SLAM uses two-dimensional (2-D) image points as inputs for 3-D location estimation. Estimation of the 3-D locations of feature points by itself is not sufficient to determine the presence of an obstacle. This paper shows how to combine a Monocular SLAM solution with image segmentation to identify an obstacle and construct a dense map around it. This paper will focus on developing an obstacle avoidance strategy for UAVs using estimates of obstacle size and location from monocular SLAM and image segmentation algorithms. The approach will be evaluated within a C/C++ and openGL based simulation environment developed specifically to investigate SLAM based obstacle avoidance techniques for UAVs. This paper will provide details of the simulation environment and also include results from this evaluation study.