A reconfigurable architecture for real-time vision systems on FPGA

This paper proposes a reconfigurable architecture for the implementation of real-time vision systems on FPGA. The architecture is composed of a library of components that can be used to perform several image processing and vision tasks. Components can be organized in one or more processing pipelines. A full frame or part of the frame can be buffered during the process to allow region-based operations that are commonly used in various vision algorithms. The functionality of the architecture is verified by implementing an algorithm for detecting ripe tomatoes in a greenhouse. The algorithm is implemented using both Matlab and the proposed architecture. An image set composed of 1511 images captured using the proposed architecture is used for testing. The results show that hardware implementation using the proposed architecture introduces a comparable accuracy with a very significant savings in processing time.