Abstract
A heterogeneous architecture offers several advantages
when designing an image processing system with real-time
constraints. We describe a platform containing a FPGA, GPU
and CPU and analyse the processing operations required in
a pedestrian detection algorithm (the Histogram of Oriented
Gradients detector) and show that real-time detection using
single and multiple accelerators with runtime switching between
several architectures is possible. When tested in a system
which incorporates communication overheads, the performance
of a heterogeneous processing system is comparable to that
of a single-accelerator system, with the limiting factor being
communications delays
when designing an image processing system with real-time
constraints. We describe a platform containing a FPGA, GPU
and CPU and analyse the processing operations required in
a pedestrian detection algorithm (the Histogram of Oriented
Gradients detector) and show that real-time detection using
single and multiple accelerators with runtime switching between
several architectures is possible. When tested in a system
which incorporates communication overheads, the performance
of a heterogeneous processing system is comparable to that
of a single-accelerator system, with the limiting factor being
communications delays
| Original language | English |
|---|---|
| Title of host publication | Proceedings of IEEE/RSJ Int. Conf. Intelligent Robots and Systems Workshops (First Workshop on Smart Cameras for Robotic Applications) |
| Publisher | IEEE |
| Publication status | Published - Oct 2012 |