Customizable FPGA-based architecture for video applications in real time

Abstract

This paper describes an efficient reconfigurable architecture suitable for full-search block-matching motion estimation with minimum latency and maximum throughput, based on a systolic array. The architecture is characterized by low memory bandwidth requirements, a modular and highly flexible structure based on Router elements to handle processing blocks interconnection from a high level pipeline scheme. Every PE in the array includes a double ALU in order to search multiple macro-blocks in parallel. The functionality has been extended to support operations involved in some low-level algorithms which include image filtering, matrix-matrix multiplication, morphological operations and pyramid processing. The architecture provides a flexible solution under real-time constraints and constitutes a platform to pursue the implementation of higher complexity algorithms. Preliminary results show that a peak performance in the order of GOPS can be achieved