TY - JOUR
T1 - The parallelization of video processing
T2 - From programming models to applications
AU - Lin, Dennis
AU - Huang, Xiaohuang
AU - Nguyen, Quang
AU - Blackburn, Joshua
AU - Rodrigues, Christopher
AU - Huang, Thomas
AU - Do, Minh N.
AU - Patel, Sanjay J.
AU - Hwu, Wen Mei W.
N1 - Funding Information:
As the hardware becomes more affordable and mass deployment of surveillance cameras becomes more common, we see increasing interest in automated methods for human activity recognition. Still, the flood of video data threatens to overwhelm even automatic systems. The TREC Video (TRECVID) Event Detection Evaluation sponsored by the National Institute of Standards and Technology illustrates the scope of the problem. In 2008, the evaluation set consisted of 50 h of data. Our sequential algorithm [2] required about 30 s/frame on a single CPU core. This meant that even running on a 16-node, 64-core cluster and processing only five frames from each second of video, we needed five days to complete the computations. This long latency posed a logistical challenge that limited our ability to experiment and tune our algorithm, and it degraded our results.
Funding Information:
We acknowledge the support of the FCRP Gigascale Systems Research Center, Intel/Microsoft Universal Parallel Computing Research Center, and VACE Program. Experiments were made possible by generous donations of hardware from NVIDIA and Intel and by NSF CNS grant 05-51665.
PY - 2009
Y1 - 2009
N2 - The explosive growth of digital video content from commodity devices and on the Internet has precipitated a renewed interest in video processing technology, which broadly encompasses the compression, enhancement, analysis, and synthesis of digital video. Video processing is computationally intensive and often has accompanying real-time or super-real-time requirements. For example, surveillance and monitoring systems need to robustly analyze video from multiple cameras in real time to automatically detect and signal unusual events. Beyond today's known applications, the continued growth of functionality and speed of video processing systems will likely further enable novel applications.
AB - The explosive growth of digital video content from commodity devices and on the Internet has precipitated a renewed interest in video processing technology, which broadly encompasses the compression, enhancement, analysis, and synthesis of digital video. Video processing is computationally intensive and often has accompanying real-time or super-real-time requirements. For example, surveillance and monitoring systems need to robustly analyze video from multiple cameras in real time to automatically detect and signal unusual events. Beyond today's known applications, the continued growth of functionality and speed of video processing systems will likely further enable novel applications.
KW - Digital video broadcasting
KW - Hardware
KW - Memory management
KW - Multicore processing
KW - Signal processing algorithms
KW - Streaming media
UR - http://www.scopus.com/inward/record.url?scp=85032752124&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85032752124&partnerID=8YFLogxK
U2 - 10.1109/MSP.2009.934116
DO - 10.1109/MSP.2009.934116
M3 - Review article
AN - SCOPUS:85032752124
SN - 1053-5888
VL - 26
SP - 103
EP - 112
JO - IEEE Signal Processing Magazine
JF - IEEE Signal Processing Magazine
IS - 6
ER -