Non-parametric belief propagation for mobile mapping sensor fusion. Issue 3 (2nd July 2016)

• Record Type: Journal Article
• Authors: Hollick, Joshua; Helmholz, Petra; Belton, David
• Is Part Of: Geo-spatial information science. Volume 19:Issue 3(2016)
• Page Start: 195
• Page End: 201
• Publication Date: 2016-07-02
• Subjects: Sensor fusion -- belief propagation -- accelerometer -- smartphone -- gyroscope -- mobile mapping; Geographic information systems -- Periodicals; Cartography -- Data processing -- Periodicals; Surveying -- Data processing -- Periodicals; Remote sensing -- Periodicals; 526.0285
• DOI: 10.1080/10095020.2016.1235816
• Languages: English
• ISSNs: 1009-5020

Abstract

Abstract: Many different forms of sensor fusion have been proposed each with its own niche. We propose a method of fusing multiple different sensor types. Our approach is built on the discrete belief propagation to fuse photogrammetry with GPS to generate three-dimensional (3D) point clouds. We propose using a non-parametric belief propagation similar to Sudderth et al's work to fuse different sensors. This technique allows continuous variables to be used, is trivially parallel making it suitable for modern many-core processors, and easily accommodates varying types and combinations of sensors. By defining the relationships between common sensors, a graph containing sensor readings can be automatically generated from sensor data without knowing a priori the availability or reliability of the sensors. This allows the use of unreliable sensors which firstly, may start and stop providing data at any time and secondly, the integration of new sensor types simply by defining their relationship with existing sensors. These features allow a flexible framework to be developed which is suitable for many tasks. Using an abstract algorithm, we can instead focus on the relationships between sensors. Where possible we use the existing relationships between sensors rather than developing new ones. These relationships are used in a belief propagation algorithm to calculate the marginal probabilities of the network. In this paper, we present the initial results from this technique and the intended course for future work.