Journals Proceedings

Abstract

Recent developments in embedded processors have enabled heterogeneous computing on mobile devices using open-access general-purpose computing languages. Following the MPEG CDVS standard, this paper presents an efficient feature computation phase, completely implemented on embedded devices supporting the OpenCL framework. Following our contribution to the MPEG-CDVS standard, we present the new born CDVS detector and its design for multicore parallel GPUs. We show how to adjust algorithmic choices and implementation details to target the intrinsic characteristics of the embedded platforms selected. We compare our GPU implementation of the ALP keypoint detector with the CPU based implementation of the CDVS standard. We present data on different GPUs showing that our solution is up to 7x faster than the CPU version. To sum up, one of the main feature of our algorithm is to be fast enough to be able to open new visual search scenarios exploiting entire real-time on-board computations with no data transfer.