EUVIMEDEUVIMED entdecken
Crossref

P‐5.11: Lightweight Hardware Accelerator for Real‐Time Object Detection on Zynq‐7020

Zhonglin Tang, Sijie Ran, Limin Yan

SID Symposium Digest of Technical Papers · 2026-08 · 57 (S1) · 1350-1354

DOI: 10.1002/sdtp.70003_357 ↗

Kurz gefasst

Fachartikel in der Zeitschrift SID Symposium Digest of Technical Papers, online erschienen am 11.09.2026. Verfasst von Zhonglin Tang und 2 weiteren Autor:innen. Die Arbeit nennt 15 Literaturangaben.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Achieving real‐time object detection on resource‐constrained embedded platforms requires a careful trade‐off among accuracy, throughput, and power. Targeting the Zynq‐7020 system‐on‐chip (SoC), this paper proposes an end‐to‐end design that combines a lightweight network structure, mixed‐precision quantization, and field‐programmable gate array (FPGA) convolution acceleration. On the algorithm side, we construct a hardware‐friendly backbone‐ neck‐head detector derived from the YOLOv4‐tiny family, including a CSPDarknet53‐tiny backbone, a lightweight FPN neck, and dual‐scale detection heads. We further rewrite operators and fuse batch normalization (BN) into convolution to improve hardware suitability. On the quantization side, we evaluate layer sensitivity using MSE and SNR and allocate 16‐bit integer (INT16) to key layers while quantizing the remaining layers to 8‐bit integer (INT8); quantization‐aware training (QAT) is applied to mitigate accuracy loss. On the hardware side, we design a tiling‐based convolution array with ping‐pong double buffering to overlap DDR transfers with computation, and use AXI Direct Memory Access (AXI‐DMA) with PS‐side scheduling for pipelined inference. On a Zynq‐7020 board with 416×416 input, the proposed system achieves 82.03 FPS with 88.96% mean average precision (mAP) and approximately 2.668 W total on‐chip power (Vivado Power), demonstrating improved real‐time performance and energy efficiency for edge vision inference.

Zitieren

APA

Tang, Z., Ran, S., & Yan, L. (2026). P‐5.11: Lightweight Hardware Accelerator for Real‐Time Object Detection on Zynq‐7020. SID Symposium Digest of Technical Papers, 57(S1), 1350-1354. https://doi.org/10.1002/sdtp.70003_357

Vancouver

Tang Z, Ran S, Yan L. P‐5.11: Lightweight Hardware Accelerator for Real‐Time Object Detection on Zynq‐7020. SID Symposium Digest of Technical Papers. 2026;57(S1):1350-1354. doi:10.1002/sdtp.70003_357

BibTeX
@article{Tang2026,
  author = {Tang, Zhonglin and Ran, Sijie and Yan, Limin},
  title = {P‐5.11: Lightweight Hardware Accelerator for Real‐Time Object Detection on Zynq‐7020},
  journal = {SID Symposium Digest of Technical Papers},
  year = {2026},
  volume = {57},
  number = {S1},
  pages = {1350--1354},
  doi = {10.1002/sdtp.70003_357}
}
Kostenlos starten →