Advanced Methods for High-Speed Template Matching Targeting FPGAs

Template matching is an important image processing algorithm for object detection and tracking tasks especially because of its very high accuracy. Yet, one major disadvantage of this algorithm is its very high computational effort due to its high complexity which results in low update rates using conventional software-based systems. Additionally, these systems are often afflicted by latency and jitter. Pre-calculating small and robust templates allows for using FPGA-based template matching as an approach to solve these problems. Therefore, the paper outlines three successive software-based approaches to find these small and robust templates. The first approach contains the basic algorithm and is refined by the following two approaches using pre-processing. The resulting templates allow for using the programmable hardware of an FPGA to cache necessary image information and, more importantly, to derive the best matching of template and source image, which results in high update rates. This paper shows novel approaches to high-speed template matching on FPGAs. The validation of these approaches has shown that the resulting tracking quality and feasibility is highly dependent on the relative size of the template in regard to the object to track. The results show tracking uncertainties between one single pixel for low and hundreds of pixels for high resolution videos.