Heterogeneous tracking system of kernel correlation filtering based on PYNQ framework

Abstract

Aiming at the complex structure and weak real-time performance of traditional visual tracking system, a scheme based on PYNQ (Python productivity for Zynq) framework is designed and deployed in Zynq heterogeneous platform. Firstly, according to the real-time requirements, the kernel-related filter tracking algorithm is selected. Then, according to the characteristics of the Zynq computing platform, the software and hardware collaborative design method is used to divide the system tasks, and the algorithm is converted into RTL (Register Transfer Level) by HLS (High Level Synthesis) development tools, and the calculation is performed. The process is optimized and exported as an IP core. Then, Python is used to import the IP core as a hardware coprocessor at the top level to implement the underlying to top-level data interaction. Finally, the system result is asynchronously updated in the Jupyter notebook. Experiments show that the system has good real-time performance, and the tracking speed averages 27.9FPS. At the same time, it has superior execution efficiency, is easy to develop and transplant, and has certain engineering reference value.