-
为了验证提出的优化算法对核相关滤波算法跟踪性能的效果,首先在上位机平台基于OTB-2015数据集进行仿真实验,上位机平台的参数为Intel i5-7300 HQ,主频2.50 GHz,软件平台为Matlab2018 a,主要采用跟踪精度以及跟踪实时性对算法的性能进行评估。
算法中的参数配置为:初始更新率$\beta $为0.02,公式(9)中历史数据的参考帧数为50,用于评估PSR以及APCE的判定系数${\delta _1}$,${\delta _2}$分别为0.25,0.30。
为了验证自适应模糊优化的可行性,对OTB-2015数据集进行了仿真实验,如图8所示,可以看出优化后的算法跟踪精度是59.74%,跟踪实时性为368 frame/s,相比于原始算法跟踪实时性仅降低了0.94%,但跟踪实时性提高了44.7%.
自适应模糊优化主要是对算法整体运算量的考量,跟踪结果的判定以及重定位则是用于提高算法的跟踪精度以及鲁棒性。下面对整体优化后的算法进行了仿真验证。
图9为不同跟踪算法在OTB-2015数据集上的跟踪精度表现,可以看出文中算法的跟踪精度为65.8%,相比传统的核相关滤波算法精度提升了8.4%,结合跟踪算法的跟踪精度以及表1中跟踪速度的表现,可以看出文中提出的算法可以在保持跟踪精度的同时,保证跟踪速度。因此,文中提出的算法有望高效地运行在嵌入式平台上。
Tracker Proposed Struct TLD CN DSST KCF Tracking speed/frame·s−1 316.8 17.4 24.4 90 19.3 254.8 Table 1. Tracking speed comparison of different tracking algorithms
为了验证文中提出的算法是否可以有效运行在嵌入式平台,结合算法的特点以及设计的需要,最终选取了融合ARM与FPGA的异构平台PYNQ-Z2完成算法的移植,并在OTB-2015[12]跟踪数据集上进行仿真实验,与原始算法的实验结果进行分析与对比,结果如表2所示。
Platform Tracking precision Tracking speed/
frame·s−1Based on ARM 60.3% 2.15 Based on ARM and FPGA 65.8% 17.28 Table 2. Comparison of tracking results based on PYNQ
Target tracking acceleration scheme adopting adaptive fuzzy optimization
doi: 10.3788/IRLA20210864
- Received Date: 2021-11-19
- Rev Recd Date: 2021-12-09
- Available Online: 2022-03-04
- Publish Date: 2022-02-28
-
Key words:
- target tracking /
- kernel correlation filtering /
- PYNQ /
- target tracking confidence /
- adaptive fuzzy
Abstract: As one of the important directions of computer vision, target tracking has a wide range of applications, such as autopilot, UAV tracking, but the target tracking algorithm cannot run effectively on embedded devices. A novel acceleration target tracking scheme based on correlation filtering was proposed to solve the problems of target tracking algorithm, such as high computation and complexity, difficulty application on the resource-constrained embedded devices. Firstly, the adaptive fuzzy algorithm was used to optimize the overall computation of the algorithm, which could decide whether to reduce the image quality based on target size. Secondly, the criterion of Peak-to-Sidelobe Rate and Average Peak-to-Correlation Energy were used to measure the reliability of tracking results, so as to realize adaptive updating of tracking model and re-search of target location. Finally, for the correlation operation and complex matrix multiplication operation in the stage of training tracking detector, which were implemented based on FPGA parallelly to improve the real-time energy efficiency of the algorithm. The proposed acceleration algorithm was deployed on PYNQ-Z2 and verified based on OTB-2015 tracking data set. The tracking accuracy and real-time performance of the algorithm were 65.8% and 17.28 frame/s, respectively, compared with the original algorithm, the tracking accuracy and real-time performance were improved by 9.12% and 703.7%, respectively.