基于可重构阵列的多粒度目标跟踪算法优化部署

Optimized Deployment of a Multi-Granularity Object Tracking Algorithm on a Reconfigurable Array

摘要:

摘要：针对目标跟踪算法在边缘计算场景下的实时性与能效比挑战，对可重构阵列的多粒度目标跟踪算法优化部署进行了研究，提高了DeepSORT目标跟踪网络卷积层硬件部署效率；对可重构阵列支持Int16/Int8/Int4三种数据粒度的计算特性进行分析，根据实际部署需求，采用动态调整数据组织方式，切换可重构阵列MAC内计算通道，优化网络参数的方法，显著提升卷积效率；实现了计算精度与执行速率之间的有效平衡；在Virtex UltraScale 440开发板上进行FPGA硬件测试，实验测试结果表明，相较于Int16模式，Int8与Int4模式在访存周期上平均降低了45.9%和72%，计算周期平均减少45.6%和73%；部署方案平均PE利用率达到95.8%，功耗仅为5.08 W，为目标跟踪算法的端侧部署提供了更灵活高效的方案。

Abstract:

Abstract: Addressing the challenges of real-time performance and energy efficiency ratio of target tracking algorithms in edge computing scenarios, this study investigates the multi-granularity target tracking algorithm optimization deployment on reconfigurable arrays, enhancing the hardware deployment efficiency of convolutional layers in the DeepSORT target tracking network. An analysis of the computational characteristics of reconfigurable arrays supporting three data granularities—Int16, Int8, and Int4—was conducted. Based on actual deployment requirements, dynamic adjustments were made to the data organization methods, switching the computational channels within the MAC units of the reconfigurable arrays and optimizing network parameters. These approaches significantly improved convolutional efficiency and achieved an effective balance between computational accuracy and execution speed. Hardware testing was performed on a Virtex UltraScale 440 FPGA development board. The experimental results show that, compared to the Int16 mode, the Int8 and Int4 modes reduced memory access cycles by an average of 45.9% and 72%, and computational cycles by an average of 45.6% and 73%, respectively. The deployment scheme achieved an average PE utilization rate of 95.8% with a power consumption of only 5.08 W, providing a more flexible and efficient solution for the edge-side deployment of target tracking algorithms.