No abstract

As one of the key steps in the processing of airborne light detection and ranging (LiDAR) data, filtering often consumes a huge amount of time and physical memory. Conventional sequential algorithms are often inefficient in filtering massive point clouds, due to their huge computational cost and Input/Output (I/O) bottlenecks. The progressive TIN (Triangulated Irregular Network) densification (PTD) filter is a commonly employed iterative method that mainly consists of the TIN generation and the judging functions. However, better quality from the progressive process comes at the cost of increasing computing time. Fortunately, it is possible to take advantage of state-of-the-art multi-core computing facilities to speed up this computationally intensive task. A streaming framework for filtering point clouds by encapsulating the PTD filter into independent computing units is proposed in this paper. Through overlapping multiple computing units and the I/O events, the efficiency of the proposed method is improved greatly. More importantly, this framework is adaptive to many filters. Experiments suggest that the proposed streaming PTD (SPTD) is able to improve the performance of massive point clouds processing and alleviate the I/O bottlenecks. The experiments also demonstrate that this SPTD allows the quick processing of massive point clouds with better adaptability. In a 12-core environment, the SPTD gains a speedup of 7.0 for filtering 249 million points.

No abstract

Abstract:In computer science, dependence analysis determines whether or not it is safe to parallelize statements in programs. In dealing with the data-intensive and computationally intensive spatial operations in processing massive volumes of geometric features, this dependence can be well utilized for exploiting the parallelism. In this paper, we propose a graph-based divide and conquer method for parallelizing spatial operations (GDCMPSO) on vector data. It can represent spatial data dependences in spatial operations through representing the vector features as graph vertices, and their computational dependences as graph edges. By this way, spatial operations can be parallelized in three steps: partitioning the graph into graph components with inter-component edges firstly, simultaneously processing multiple subtasks indicated by the graph components secondly and finally handling remainder tasks denoted by the inter-component edges. To demonstrate how it works, buffer operation and intersection operation under this paradigm are conducted. In a 12-core environment, the two spatial operations both gain obvious performance improvements, and the speedups are more than eight. The testing results suggest that GDCMPSO contributes to a method for parallelizing spatial operations and can greatly improve the computing efficiency on multi-core architectures.