A model-based design flow for embedded vision applications on heterogeneous architectures

The ability to gather information from images is straightforward to human, and one of the principal input to understand external world. Computer vision (CV) is the process to extract such knowledge from the visual domain in an algorithmic fashion. The requested computational power to process these information is very high. Until recently, the only feasible way to meet non-functional requirements like performance was to develop custom hardware, which is costly, time-consuming and can not be reused in a general purpose. The recent introduction of low-power and low-cost heterogeneous embedded boards, in which CPUs are combine with heterogeneous accelerators like GPUs, DSPs and FPGAs, can combine the hardware efficiency needed for non-functional requirements with the flexibility of software development. Embedded vision is the term used to identify the application of the aforementioned CV algorithms applied in the embedded field, which usually requires to satisfy, other than functional requirements, also non-functional requirements such as real-time performance, power, and energy efficiency. Rapid prototyping, early algorithm parametrization, testing, and validation of complex embedded video applications for such heterogeneous architectures is a very challenging task. This thesis presents a comprehensive framework that: 1) Is based on a model-based paradigm. Differently from the standard approaches at the state of the art that require designers to manually model the algorithm in any programming language, the proposed approach allows for a rapid prototyping, algorithm validation and parametrization in a model-based design environment (i.e., Matlab/Simulink). The framework relies on a multi-level design and verification flow by which the high-level model is then semi-automatically refined towards the final automatic synthesis into the target hardware device. 2) Relies on a polyglot parallel programming model. The proposed model combines different programming languages and environments such as C/C++, OpenMP, PThreads, OpenVX, OpenCV, and CUDA to best exploit different levels of parallelism while guaranteeing a semi-automatic customization. 3) Optimizes the application performance and energy efficiency through a novel algorithm for the mapping and scheduling of the application 3 tasks on the heterogeneous computing elements of the device. Such an algorithm, called exclusive earliest finish time (XEFT), takes into consideration the possible multiple implementation of tasks for different computing elements (e.g., a task primitive for CPU and an equivalent parallel implementation for GPU). It introduces and takes advantage of the notion of exclusive overlap between primitives to improve the load balancing. This thesis is the result of three years of research activity, during which all the incremental steps made to compose the framework have been tested on real case studies