A GPU-Accelerated Approach for Feature Tracking in Time-Varying Imagery Datasets.

Area of Science:

• Computer Science
• Image Processing
• Big Data Analytics

Background:

• Large-scale, time-varying imagery datasets pose significant computational challenges.
• Existing methods struggle with memory limitations and data dependencies.

Purpose of the Study:

• To develop a novel parallel connected component labeling (CCL) algorithm for efficient feature detection and tracking in large, time-varying imagery.
• To address the out-of-core data management problem for massive datasets on GPU architectures.

Main Methods:

• A parallel CCL algorithm tailored for GPU architectures, achieving pixel-level parallelism.
• Out-of-core data management involving streaming, partitioning, and batch processing of image frames.
• A novel parallel batch merging algorithm to reconnect feature regions separated by partitioning.

Main Results:

• Efficient processing of terabyte-scale, time-varying radar weather datasets on a single GPU.
• Demonstrated performance advantages over an efficient CPU cluster implementation.
• The algorithm successfully detects and tracks feature regions in large imagery.

Conclusions:

• The proposed parallel CCL algorithm with out-of-core data management offers an efficient solution for big data image analysis on GPUs.
• This approach enables effective processing of massive, time-varying datasets previously intractable for single workstations.
• The method provides a scalable and performant alternative to traditional CPU cluster implementations for scientific applications.