Investigation of Multicore SoCs for On-Board Feature Detection and Segmentation of Images

Abstract

Increased resolution and bit-depth of image sensors present many challenges in feature-detection and segmentation algorithms for on-board image processing. This paper provides new optimizations, results, and analysis of two such algorithms on multicore, system-on-chip (SoC) devices including the DSP-based TI 66AK2H12 SoC and the GPU-based Nvidia Tegra K1/X1/X2 SoC. We chose FAST (Features from Accelerated Segment Test) as our feature-detection algorithm and GMS (Gaussian mean-shift) as our segmentation algorithm. Our optimized designs for the FAST algorithm achieves real-time performance on all the SoCs at low power. Furthermore, we compare and analyze the performance-per-watt of a compute-intensive GMS segmentation algorithm on the SoCs.