Out-of-core GPU accelerated surface reconstruction for large industrial environment monitoring

Abstract

A parallel implementation of a surface reconstruction algorithm is presented. This algorithm uses the vector field surface representation and was adapted in a previous work by the authors to handle large scale environment reconstruction. Two parallel implementations with different memory requirements and processing speeds are described and compared. These parallel implementations increase the vector field computation speed by a factor of up to 31 times relative to a purely serial implementation. The method is demonstrated on different datasets captured on the sites of Hydro-Quebec using a variety of sensors: LiDAR, sonar and the WireScan, an underwater laser scanner designed at our laboratory.