Dajun Jin: Implementation of Feature Detection in High-Level Synthesis for HiFlipVX (project work)

28.09.2020,  2.00 pm

Invitation to the presentation of Dajun Jin

Topic: Implementation of Feature Detection in High-Level Synthesis for HiFlipVX

Project: Project work

Supervisor: Lester Kalms

Abstract: In this work the student should implement the ORB feature detection algorithm [1] using High-Level Synthesis (HLS) for FPGAs. A hardware implementation of this algorithm exist in [2]. Figure 0.1 shows a simple block diagram of the algorithm. The student should implement the missing functions based on the OpenVX standard [3]. The functions of the student will be added to the HiFlipVX library [4, 5]. The function should be the baseline of how to write image processing functions in HLS using C++ templates. Some of the functions of the ORB feature detector already exist in this library. Existing functions are: Color Conversion, Non-Maxima Suppression and Scale. Missing functions are: Fast Corners [6] and Retain Best. For training HLS the student should implement the Control Flow function described in the OpenVX standard. Using the implemented functions and the missing functions, the ORB detector should be build and tested. The implementation should be comparable in performance and resource utilization to [2]. The hardware implementation should also be tested in terms of accuracy (repeatability). This can be done by comparing the results that the function produces with the OpenCV implementation of the ORB. As optional task the student can implement the canny edge detector using the HiFlipVX library and evaluate the implementation.