An FPGA-based accelerator for multiple real-time template matching

Object tracking with multiple dense templates is a challenging problem in the context of continuous monitoring video cameras. Applications of the multiple match template technique range from tracking of multiple independent objects to several pose variations of a single object. Due to its high accuracy and tolerance to brightness and contrast changes, the zero mean normalized cross-correlation(ZNCC) was selected as similarity measure. This paper proposes an FPGA architecture that explores a full pipeline implementation and maximizes the internal data reuse to calculate several ZNCC-based template matching in an efficient approach. Experimental results shows that the proposed implementation achieves the real-time performance reaching up to 30fps running ten parallel ZNCCs in a single Stratix IV FPGA.