Scholarlink

The increasing demand for real-time, energy-efficient perception in autonomous navigation systems has amplified the need for lightweight yet accurate object detection frameworks. Traditional deep convolutional models such as YOLOv5 deliver high detection performance but remain computationally intensive, limiting their applicability on edge devices with restricted power and latency budgets. Conversely, Spiking Neural Networks (SNNs) offer biologically inspired, event-driven computation with significantly lower energy requirements, though they often fall short in detection precision and robustness when used in isolation. This paper introduces a novel hybrid obstacle detection pipeline that integrates the spatial accuracy of YOLOv5 with the energy efficiency of a Leaky Integrate-and-Fire (LIF) based SNN classifier. Two architectures are explored: a modular hybrid pipeline where YOLOv5 performs detection and SNN handles classification of cropped ROI [Revised: ROI (Region of Interest)]s, and a customized and integrated YOLO → SNN architecture wherein bounding box-aligned feature maps are directly encoded and classified by the SNN using surrogate gradient optimization. Both systems decouple localization and classification to reduce computational redundancy and enable deployment flexibility across resource-constrained platforms. Extensive experiments conducted on a self-driving car dataset comprising five key object categories validate the effectiveness of the proposed architectures. The customized and integrated YOLO → SNN system achieves a mean Average Precision (mAP@0.5) of 0.79 while maintaining 68 + FPS [Revised: FPS (Frames per Second)], outperforming standalone and hybrid baselines in both energy efficiency and classification robustness. Comparative quantitative and qualitative evaluations demonstrate that the inclusion of SNN-based classification not only reduces computational overhead but also enhances interpretability and edge deployment potential, positioning the customized and integrated hybrid model as a scalable, computation -friendly solution for real-time obstacle detection.

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

Neuromorphic Computing Environment for Optimized Real-Time Obstacle Detection Using Integrated & Hybrid YOLO-SNN Models with Advanced Simulation

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • R. Vinay,
  • K. Pradeep Kumar,
  • Sugandha Saxena,
  • M. Lakshmanan,
  • B. N. Aryalekshmi,
  • Priya Singh,
  • Salma Itagi

摘要

The increasing demand for real-time, energy-efficient perception in autonomous navigation systems has amplified the need for lightweight yet accurate object detection frameworks. Traditional deep convolutional models such as YOLOv5 deliver high detection performance but remain computationally intensive, limiting their applicability on edge devices with restricted power and latency budgets. Conversely, Spiking Neural Networks (SNNs) offer biologically inspired, event-driven computation with significantly lower energy requirements, though they often fall short in detection precision and robustness when used in isolation. This paper introduces a novel hybrid obstacle detection pipeline that integrates the spatial accuracy of YOLOv5 with the energy efficiency of a Leaky Integrate-and-Fire (LIF) based SNN classifier. Two architectures are explored: a modular hybrid pipeline where YOLOv5 performs detection and SNN handles classification of cropped ROI [Revised: ROI (Region of Interest)]s, and a customized and integrated YOLO → SNN architecture wherein bounding box-aligned feature maps are directly encoded and classified by the SNN using surrogate gradient optimization. Both systems decouple localization and classification to reduce computational redundancy and enable deployment flexibility across resource-constrained platforms. Extensive experiments conducted on a self-driving car dataset comprising five key object categories validate the effectiveness of the proposed architectures. The customized and integrated YOLO → SNN system achieves a mean Average Precision (mAP@0.5) of 0.79 while maintaining 68 + FPS [Revised: FPS (Frames per Second)], outperforming standalone and hybrid baselines in both energy efficiency and classification robustness. Comparative quantitative and qualitative evaluations demonstrate that the inclusion of SNN-based classification not only reduces computational overhead but also enhances interpretability and edge deployment potential, positioning the customized and integrated hybrid model as a scalable, computation -friendly solution for real-time obstacle detection.