A Review of Design Methodologies and Evaluation Techniques for FPGA-Based Visual Object Tracking Systems

Abstract

In recent years, computer vision algorithms have improved from conventional image processing to deep learning approaches. Meanwhile, complex but flexible FPGA-based platforms have made possible the development of challenging real-time heterogeneous systems for visual object tracking (VOT). This study presents a comprehensive review of design methodologies, algorithms, and evaluation techniques of 21 FPGA-based VOT systems reported in literature from 2017 to 2023. Five design methodology categories are described: Region matching, Feature matching, Machine learning, Deep learning, and Hybrid systems. FPGA, SoC-FPGA and MPSoC platforms for VOT system implementations are considered. Relevant evaluation techniques and metrics are reviewed as part of dataset benchmarks or toy datasets. In order to propose an insight of the FPGA-based VOT systems, each topic presents their comparative analysis and discussion. Finally, main conclusions, recommendations and perspectives are presented.