Scholarlink

In recent years, neuromorphic technologies have gained increasing significance in the development of autonomous robotic systems and human-robot interaction systems. The relevance of this field is driven by the reduction in computational complexity and energy consumption through mechanisms that mimic the principles of biological systems. The article explores the application of spiking neural networks (SNN) for object detection using dynamic vision sensor (DVS) cameras, which are characterized by a high dynamic range and speed. The potential for improving accuracy in static scenes through multimodal data processing from DVS and RGB cameras is also examined. The research primarily focuses on using a training method for sequences of variable length, which reduces prediction latency and enables the model to be trained for streaming data processing. The effectiveness of the proposed approaches is analyzed on the Gen 1 and DSEC-Detection datasets, demonstrating significant improvements in prediction accuracy and energy efficiency. The results confirm the practical applicability of the proposed SNN training methods and their potential to enhance the accuracy and energy efficiency of autonomous robotic systems. Code: https://github.com/KirillHit/twl_spike_yolo .

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

Using Spiking Neural Networks for Event and Multimodal Data Processing in Object Detection Tasks

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Kirill S. Khitushkin,
  • Tim T. Isakov,
  • Aleksandr V. Bakhshiev

摘要

In recent years, neuromorphic technologies have gained increasing significance in the development of autonomous robotic systems and human-robot interaction systems. The relevance of this field is driven by the reduction in computational complexity and energy consumption through mechanisms that mimic the principles of biological systems. The article explores the application of spiking neural networks (SNN) for object detection using dynamic vision sensor (DVS) cameras, which are characterized by a high dynamic range and speed. The potential for improving accuracy in static scenes through multimodal data processing from DVS and RGB cameras is also examined. The research primarily focuses on using a training method for sequences of variable length, which reduces prediction latency and enables the model to be trained for streaming data processing. The effectiveness of the proposed approaches is analyzed on the Gen 1 and DSEC-Detection datasets, demonstrating significant improvements in prediction accuracy and energy efficiency. The results confirm the practical applicability of the proposed SNN training methods and their potential to enhance the accuracy and energy efficiency of autonomous robotic systems. Code: https://github.com/KirillHit/twl_spike_yolo .