실시간 물체 인식을 위한 특징 추출 알고리즘의 병렬화

초록

This paper presents a fast feature extraction method for mobile robots by parallel processing based on OpenMP and SSE (Streaming SIMD Extension) programming. Recognizing scenes or objects in real-time is a difficult task, due to the variety of features. To complete this task, a parallel feature extraction detector and descriptor via OpenMP and SSE is proposed. In the first step, the algorithms and codes are optimized, in order to be implemented by parallel processing. The implemented parallel algorithms are debugged, to maintain the same level of performance. The process of extracting key points and obtaining the dominant orientation with respect to key points is parallelized via these steps. After extraction, a parallel descriptor via SSE instructions is constructed. As a result, the proposed Parallel-Hessian accelerates processing by up to two and half times over previous detectors, and the descriptor achieves an acceleration of up to four and half times compared to the original SIFT, while maintaining a robust performance measured via a standard evaluation set.