The CACTuS multi-object visual tracking algorithm on a heterogeneous computing system

Multi-core CPUs, Graphics Processing Units (GPUs) and Field Programmable Gate Arrays (FPGAs) are increasingly being combined into heterogeneous computing systems to accelerate the performance of image processing applications. In the realm of image processing, heterogeneous systems that combine computing platforms with different computing characteristics are an attractive approach because they give opportunity for various forms of parallelism that exist within complex image processing applications to be exploited in a multitude of ways. In this paper we detail our efforts to develop implementations of the Competitive Attentional Correlation Tracker using Shape (CACTuS) multi-object visual tracking algorithm for a heterogeneous system comprised of a multi-core CPU, GPU and FPGA. We discuss the development of modular components of this algorithm for each of these platforms, and explore how we used detailed performance measures to guide the integration of components from the various platforms together to create a heterogeneous version of the algorithm that utilises all three disparate platforms.