Previous Abstract Return to Session C1 Next Abstract

Session C1: Navigation Using Environmental Features

Compressed Smoothing of Pseudo-SLAM for UAV Navigation Applications
Jonghyuk Kim, Centre for Autonomous Systems, University of Technology, Australia; Jose Guivant, University of New South Wales, Australia
Location: Atrium Ballroom
Alternate Number 4

This paper addresses the fusion of the pseudorange/pseudorange rate observations from global navigation satellite system (GNSS), and the inertial-visual simultaneous localization and mapping (SLAM) to achieve reliable navigation of unmanned aerial vehicles (UAVs). This work extends the previous work on Compressed Pseudo-SLAM which is an all-source-based navigation framework, integrating GNSS pseudorange/carrier-phase measurements, inertial measurement, and visual features in a computationally efficient way. We proposes a novel compressed smoothing algorithm by developing a fixed-region smoother in which the smoothing region is dynamically defined around the vehicle, while accumulating (or compressing) the locally acquired information to reduce the computational complexity of the SLAM system. This method enables the local information compression and a much lower frequency of switching events (or marginalization), thus offering the computational benefit.
A real-flight dataset collected from a fixed-wing UAV platform (a push-prop type platform with 2.9m wingspan and conventional tricycle undercarriage) will be used to demonstrate the methods. It will show the improved performance of the smoothing version of the compressed pseudo-SLAM compared to the previous filtering version, whilst achieving significant computational benefit. It will also demonstrate that the system can navigate effectively under the condition of one satellite and one landmark observations. A total of 85 landmarks are registered during the flight, and a 17 vehicle-state smoother is implemented, resulting in the covariance matrix size of 272x272. The computational cost will be analyzed, showing less than 1.5ms per update cycle, thanks to the compressed smoothing, and thus demonstrating the real-time efficiency of the method.



Previous Abstract Return to Session C1 Next Abstract