Scholarlink

<p>Road accidents are one of the major contributors to mortality in India with nearly 1.68 lakh deaths in 2023 accounting for nearly 19 deaths per hour. While road accidents aren’t always “<i>fatal</i>”, delayed detection leads to the loss of many lives. Deep Learning deliberately showed terrific results for a lot of tasks, but for a long time, the Deep Learning models utilized the standard Multi-Layered Perceptron (<Emphasis FontCategory="NonProportional">MLP</Emphasis>) architecture as its’ skeleton. Recently, a state-of-the-art architecture—Kolmogorov Arnold Network (<Emphasis FontCategory="NonProportional">KAN</Emphasis>) showed up, with variable activation functions in each edge of the graphical representation of the network. Several works have inherited the <Emphasis FontCategory="NonProportional">KAN</Emphasis> architecture and have shown better results in many instances. This research proposes a class of accident detection models, SADAK (<b>S</b>imple, and <b>A</b>utomatic <b>D</b>etection of <b>A</b>ccidents on roads using <b>K</b>olmogorov–Arnold Networks) for detection of road accidents inheriting <Emphasis FontCategory="NonProportional">KAN</Emphasis> architecture. To benchmark its efficiency, several well-known <Emphasis FontCategory="NonProportional">MLP</Emphasis> based architectures have been considered, alongside some state-of-the-art models, which result in accuracies as high as 89% using the <Emphasis FontCategory="NonProportional">MLP</Emphasis> (<Emphasis FontCategory="NonProportional">Vision Transformer</Emphasis>), and 97% using the <Emphasis FontCategory="NonProportional">KAN</Emphasis> architecture (<Emphasis FontCategory="NonProportional">KAN-Convolutional-MLP</Emphasis>).</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

SADAK: Simple, and Automatic Detection of Accidents on roads using Kolmogorov-Arnold Networks

  • 期刊论文

  • 发表日期：

    2025年1月28日

查看全文
  • Anurag Dutta,
  • Pijush Kanti Kumar,
  • K. Lakshmanan,
  • Soubhik Ghosh

摘要

Road accidents are one of the major contributors to mortality in India with nearly 1.68 lakh deaths in 2023 accounting for nearly 19 deaths per hour. While road accidents aren’t always “fatal”, delayed detection leads to the loss of many lives. Deep Learning deliberately showed terrific results for a lot of tasks, but for a long time, the Deep Learning models utilized the standard Multi-Layered Perceptron (MLP) architecture as its’ skeleton. Recently, a state-of-the-art architecture—Kolmogorov Arnold Network (KAN) showed up, with variable activation functions in each edge of the graphical representation of the network. Several works have inherited the KAN architecture and have shown better results in many instances. This research proposes a class of accident detection models, SADAK (Simple, and Automatic Detection of Accidents on roads using Kolmogorov–Arnold Networks) for detection of road accidents inheriting KAN architecture. To benchmark its efficiency, several well-known MLP based architectures have been considered, alongside some state-of-the-art models, which result in accuracies as high as 89% using the MLP (Vision Transformer), and 97% using the KAN architecture (KAN-Convolutional-MLP).