Run-Processing: A Coherence-oriented Processing Method and its Hardware Architecture for Real-time Video Object Segmentation

A coherence-oriented run-processing method and its hardware architecture is proposed for multiple widely-used techniques in video object segmentation algorithms including connected component labeling, binary morphology, seeded region growing and bitwise operations of binary masks. Experimented result shows that this run-domain processing method allows significant computation acceleration, memory reduction and bus-bandwidth reduction as well in each operation with limited resources. Via hardware acceleration, a shadow-cancellation video object segmentation algorithm achieves 151fps on D1 (720x480) video. Run-processing reduces temporary data storage by 78% and takes 13.9% computation cycles compared to conventional pixel-based methods.