Parallelization Of Object-oriented Machine Vision Algorithms For Embedded GPUs

Decision logic based on machine vision is among the prevalent building blocks of today's smart devices, vehicles and things. A vast community in both industry and academia works to design and improve machine vision algorithms. Design phases often disregard end device architecture and specifics, relying on pseudo-code, scripting or object oriented approaches. Object-oriented code increases the complexity of porting and optimization for end multi-processor systems. Parallelization becomes required to reach real-time performance, with GPUs being among the most frequently utilized accelerators. In this paper, we propose methods and steps to efficiently parallelize machine vision code designed in object-oriented paradigm. We address challenges such as call stack usage optimization, code compacting, object dereferencing and partitioning for efficient execution on GPU kernels. Finally, we evaluate the proposed method by porting and optimizing an automotive camera-based vehicle detection algorithm.