Scholarlink

<p>With the widespread application of autonomous underwater vehicles (AUVs) in marine resource exploration, environmental monitoring, and underwater operations, their onboard perception systems face multiple challenges in detection accuracy, real-time performance, and energy efficiency in complex underwater environments. Traditional GPU-based target detection solutions are challenging to meet the actual deployment requirements in AUVs due to their high power consumption and low portability. To this end, this paper proposes an efficient target detection deployment solution with collaborative software and hardware optimization. At the algorithm level, the YOLOX network is improved, the lightweight MobileVitBlock module is integrated, which reduces computational complexity, and the hardware-friendly structure is optimized; at the hardware architecture level, a dedicated three-dimensional systolic array architecture is designed for the algorithm, which boosts parallel efficiency. It supports dual computing modes of convolution and group attention, and achieves efficient execution of hybrid operators through a unified computing engine. the collaborative design solution was deployed on VU9P, reaching a peak throughput of 608 GOP/s and an average energy efficiency of 69.09 GOP/W, 3.72 times higher than GPU and about 13 times higher than CPU. The inference frame rate can reach 51 FPS, exhibiting real-time inference capability and rapid response characteristics, and showing potential for application in resource-constrained underwater intelligent platforms such as AUVs.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

FPGA-based real-time underwater object detection system for AUVs

  • 期刊论文

  • 发表日期：

    2025年10月24日

查看全文
  • Zhenbo Qi,
  • Qingzeng Song,
  • Yongjiang Xue,
  • Qiuyan Wang,
  • Fei Qiao

摘要

With the widespread application of autonomous underwater vehicles (AUVs) in marine resource exploration, environmental monitoring, and underwater operations, their onboard perception systems face multiple challenges in detection accuracy, real-time performance, and energy efficiency in complex underwater environments. Traditional GPU-based target detection solutions are challenging to meet the actual deployment requirements in AUVs due to their high power consumption and low portability. To this end, this paper proposes an efficient target detection deployment solution with collaborative software and hardware optimization. At the algorithm level, the YOLOX network is improved, the lightweight MobileVitBlock module is integrated, which reduces computational complexity, and the hardware-friendly structure is optimized; at the hardware architecture level, a dedicated three-dimensional systolic array architecture is designed for the algorithm, which boosts parallel efficiency. It supports dual computing modes of convolution and group attention, and achieves efficient execution of hybrid operators through a unified computing engine. the collaborative design solution was deployed on VU9P, reaching a peak throughput of 608 GOP/s and an average energy efficiency of 69.09 GOP/W, 3.72 times higher than GPU and about 13 times higher than CPU. The inference frame rate can reach 51 FPS, exhibiting real-time inference capability and rapid response characteristics, and showing potential for application in resource-constrained underwater intelligent platforms such as AUVs.