Multiscale Architectures and Parallel Algorithms for Video Object Tracking
Abstract
Implementation and performance of an extended set of parallel multicore video processing chain of modules were investigated including Cell/B.E. flux tensor for motion detection, morphology, connected component labeling, and object statistics for blob extraction. A limitation of earlier work included unsatisfactory end-to-end performance and a need for better integration within the Net-Centric Exploitation and Tracking (N-CET) software framework. We examined software integration for Phoenix support, reducing thread overhead, video processing module interoperability, enhancing streaming performance, evaluating power requirement tradeoffs between different algorithms, assisting with tighter integration of the individual modules and gathering benchmarking data to analyze performance bottlenecks in communication pathways and computational algorithms. The video processing modules were ported to the IBM QS20/QS22 Blade architectures with 16 Synergistic Processing Elements for improved numerical computation performance, especially for the flux tensor computation.
Document Details
- Document Type
- Technical Report
- Publication Date
- Oct 01, 2011
- Accession Number
- ADA550808
Entities
People
- Kannappan Palaniappan
Organizations
- University of Missouri