A Hybrid Multi-core Architecture for Real-time Video Tracking

In this paper, we present an implementation of real-time video tracking on a novel reconfigurable multi-core architecture capable of reacting to changing workload while minimizing the number of active cores. The system is comprised of multiple processor cores executing sequential software threads, and hardware cores implemented in an FPGA executing dynamically reconfigurable hardware threads. SW and HW threads interact using a unified multithreaded programming model, which allows on-the-fly reconfiguration to shift workload between hardware and software components. Our self-adaptation technique effectively re-partitions threads across hardware and software cores to keep the performance of a video object tracking application within a predefined budget while minimizing the number of used processing elements and, thus, saving power consumption.