Scholarlink

<p>Biological visual systems naturally include multiple visual motion detection mechanisms, of which some intrinsic motion-sensitive neurons can perceive and track changes of multi-object motion. Nevertheless, it remains open whether such mechanisms can be extracted to construct multi-object tracking models. To fill this gap, a feedforward crab visual neural network with sixteen subnetworks is, for the first time, developed to detect changes of motion targets in panoramic scenes, inspired by the specific response characteristics of monostratified lobula giant neurons. Thereafter, an artificial multi-object tracking system with the neural network is developed to achieve omnidirectional target motion tracking, in which the properties of the crab’s centrifugal neurons are borrowed to construct two computational models, to compute the match degree between targets in adjacent frames and recognize the label of each target in the field of view. The theoretical analysis indicates that the computational efficiency of the target tracking system with <i>O</i>(222<i>MN</i>) greatly depends on the input image’s resolution. Two experimental conclusions can be clearly drawn: (i) the neural network can well capture visual motion characteristics of all the motion targets in the panoramic scene; and (ii) the target tracking system can significantly outperform the compared models with a high precision of target tracking, while only taking about 0.048&#xa0;s to execute on-line multi-object tracking for each frame with resolution 120 × 120.</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

Omnidirectional motion perception inspired artificial crab visual neural network and related multi-object tracking

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月15日

查看全文
  • Chenghan Liu,
  • Zhuhong Zhang

摘要

Biological visual systems naturally include multiple visual motion detection mechanisms, of which some intrinsic motion-sensitive neurons can perceive and track changes of multi-object motion. Nevertheless, it remains open whether such mechanisms can be extracted to construct multi-object tracking models. To fill this gap, a feedforward crab visual neural network with sixteen subnetworks is, for the first time, developed to detect changes of motion targets in panoramic scenes, inspired by the specific response characteristics of monostratified lobula giant neurons. Thereafter, an artificial multi-object tracking system with the neural network is developed to achieve omnidirectional target motion tracking, in which the properties of the crab’s centrifugal neurons are borrowed to construct two computational models, to compute the match degree between targets in adjacent frames and recognize the label of each target in the field of view. The theoretical analysis indicates that the computational efficiency of the target tracking system with O(222MN) greatly depends on the input image’s resolution. Two experimental conclusions can be clearly drawn: (i) the neural network can well capture visual motion characteristics of all the motion targets in the panoramic scene; and (ii) the target tracking system can significantly outperform the compared models with a high precision of target tracking, while only taking about 0.048 s to execute on-line multi-object tracking for each frame with resolution 120 × 120.