A Simultaneous Localization and 3D Mapping (SLAM) System using Airborne Cameras
Recently, there has been an increasing need for on-line 3D environment reconstruction methods in a wide range of applications, like robot navigation and mapping, and augmented and virtual reality. In order to implement this, the concept of Simultaneous Localization and Mapping(SLAM) has been proposed, which refers to the solution of keeping track of an agent relative to its environment and in the meantime, building a 3D model of the environment.
Considering huge advantages in cost and accessibility, visual sensors, like color camera, are widely employed by current SLAM systems, i.e. visual SLAM(vSLAM). However, most current vSLAM systems have used traditional image descriptors for pose estimation, like ORB-SLAM using Oriented FAST and Rotated BRIEF(ORB) descriptors, which often do not obtain good results when running on scenes without rich textures. Besides, a dense/semi-dense, high-precision 3D environment reconstruction is usually quite difficult to be achieved by already existing SLAM solutions.
In contrast to traditional indirect (based on features) vSLAM solutions, this research project is aimed to develop a direct vSLAM system, which allows to build at least semi-dense, consistent 3D models of the environment.
In this project, we use the colored data of images directly to achieve direct image alignment and to obtain the relative transformation, without implementing conventional time-consuming features extraction and description processes. Based on an estimated poses, the scene depths can be recovered and updated. Then, a sliding-window optimization strategy will be employed, where a local state information matrix will be calculated and the estimated errors can be further minimized. Finally, if a dense, high-precision 3D reconstruction is required, a standard global optimization method, like Bundle Adjustment (BA), can be applied.
The experimental setup consists of a unmanned aerial vehicle (UAV), an on-board mini-PC, a computer on the ground and a time-of-flight (tof) camera. In order to avoid the scale-drift problem in monocular SLAM system, we employ a tof camera as the visual sensor, which can provide coarse depth measurements as initial values. To keep this system running in real time, the algorithm can be dived into two parts: pose estimation and mapping algorithm on mini-PC, and optimization algorithm on ground PC. The data transmission between on-board PC and ground PC can be achieved using a data link.