Abstract:

In order to solve the problems of feature extraction in deep convolutional features visual tracking algorithm, such as large amount of computation, slow speed and difficulty in application on embedded platform, a visual tracking scheme based on PYNQ framework is proposed and deployed on Zynq heterogeneous platform. Firstly, a visual tracking algorithm based on deep convolutional features is designed. Then, according to the characteristics of the algorithm, the software and hardware are divided to complete the system-on-chip construction. Then, the calculation process of deep convolutional features extraction is optimized in parallel and derived as an accelerated IP core. Finally, via Jupyter Notebooks in the PYNQ framework, the accelerated IP core can be used as a hardware coprocessor to achieve data interaction from the bottom to the top. The experimental results show that the algorithm achieves good tracking accuracy on the benchmark OTB-2015 and UAV123. The tracking speed is up to 30 times higher than when the accelerated IP core is not integrated. Under the circumstance of considering the tracking robustness, the heterogeneous tracking system has high execution efficiency, good portability and engineering application value.

Key words: PYNQ framework, object tracking, deep convolutional features, Zynq, accelerate

摘要：

针对深度卷积特征目标跟踪算法中特征提取计算量大、速度慢、难以在嵌入式平台上应用的问题，提出了一种基于PYNQ框架的目标跟踪方案，并将其部署在Zynq异构平台。首先设计基于深度卷积特征的目标跟踪算法；根据算法的特点进行软硬件划分，完成片上系统的构建；然后针对深度卷积特征提取的计算过程进行并行优化，导出加速IP核；最后在PYNQ框架中通过Jupyter Notebooks，使用Python语言调用加速IP核作为硬件协处理器，实现底层到顶层的数据交互。实验结果表明，算法在通用数据集OTB-2015、UAV123上取得了良好的跟踪精度；跟踪速度与未集成加速IP核时相比，提升可达30倍。在兼顾跟踪稳健性的情况下，异构跟踪系统执行效率高，可移植性好，具有工程应用价值。

关键词: PYNQ框架, 目标跟踪, 深度卷积特征, Zynq, 加速