Optimising Multiprocessor Image-Based Control Through Pipelining and Parallelism

Abstract

Image-based control (IBC) systems have a long sensing delay due to compute-intensive image processing. Modern multiprocessor IBC implementations consider either parallelisation of the sensing task or pipelining of the control loop to cope with this long delay. However, the impact of both parallelisation and pipelining together on the quality-of-control (QoC) of IBC systems is not explored in the literature. We present a model-based design method for multiprocessor IBC implementation, considering both parallelisation and pipelining together. In particular, we address the following problem: For a given platform allocation, what is the optimal degree of pipelining and degree of parallelisation required to maximise the QoC? The proposed method takes into account image-workload variations, inter-frame dependencies and platform constraints. The application is efficiently modelled and analysed using a scenario-aware dataflow graph, and an implementation-aware switched controller is designed that optimises QoC and guarantees stability. We validate the proposed method using simulations and hardware-in-the-loop experiments, considering a lane-keeping assist system.