Porównanie sprzętowych implementacji dwóch popularnych detektorów cech punktowych

W artykule zaprezentowano sprzętowe implementacje dwóch detektorów narożników - detektora Harrisa i detektora FAST - w strukturach FPGA. Prędkość przetwarzania nie ustępuje prędkości uzyskiwanej na współczesnych komputerach osobistych, jednakże zastosowanie niedrogich układów FPGA umożliwia ograniczenie poboru mocy, a także kosztu oraz wymiarów kompletnego systemu. W artykule zawarto opis obu algorytmów, schematy blokowe ich sprzętowych implementacji, a także podsumowanie i porównanie ilości zasobów układu FPGA wykorzystywanych przez obie implementacje. Wykonano również wstępną analizę wyników uzyskanych przez zastosowanie zaimplementowanych detektorów na sekwencji obrazów.

Many contemporary computer and machine vision applications require finding corresponding points in image sequences. For that purpose many point feature detectors have been developed. Most of them detect corners, i.e. points that mark object boundaries, or boundaries of significant object parts as features. In this paper there are presented the implementations of two popular corner detectors - the Harris [2] and FAST [3] corner detector - in FPGA structure. The proposed solutions enable processing of 512x512 pixel, 8-bit grayscale image data with the speed of over 400 frames per second (FAST), and over 350 frames per second (Harris). The processing speeds are the same or even better than those that can be achieved using modern high-performance PCs. FPGA implementations, however, are less power-hungry, relatively inexpensive and more compact, which is critical in many applications. Our implementations are targeted at applications in mobile robotics. The paper contains a short description of the implemented algorithms, block diagrams of the implemented architectures, as well as the summary of the FPGA resources required by both implementations. A preliminary analysis of performance of the implemented algorithms with regards to feature repeatability is also carried out. The results show that the implementation of the FAST algorithm has better performance in terms of speed. Also, the FAST algorithm performs better on image sequences with strong structure - urban, interiors etc. The Harris detector implementation, although in general slower and a little more resource-hungry than the FAST implementation (requires hardware multipliers), demonstrates better performance on poorly structured scene sequences - grass, dirt roads etc. These conclusions are consistent with the results of research carried out before [3, 4].