跳转至

文章背景与核心概要

两轮机动车(PTW)驾驶员在复杂的道路交通环境中面临着极高的安全风险,而诸如时间压力(TP)等认知应激因素会显著加剧这一威胁。为了解决传统碰撞风险评估模型计算开销大、难以在资源受限的边缘设备上部署以及忽略认知压力影响的问题,研究人员开发了 MotoSafety 架构。该研究通过引入独特的“学习型时间重要性”原则,并结合大规模模拟驾驶数据集,在保证极高分类与预测准确率的同时,大幅降低了计算复杂度和推理延迟。

MotoSafety 不仅在碰撞风险分类和时间序列预测上超越了现有的多种基线模型(如 TimesNet、LLM4TS、Time-LLM 和 iTransformer),还展现出了极高的实用性和跨领域迁移能力,仅需少量的传感器特征即可实现低成本、高效率的边缘端实时部署,为未来的智能交通安全系统提供了全新的技术支撑。


MotoSafety: Edge-AI with Learned Temporal Importance for Two-Wheeler Collision Risk Assessment Under Time Pressure

MotoSafety: Edge-AI with Learned Temporal Importance for Two-Wheeler Collision Risk Assessment Under Time Pressure

arXiv: 2608.17823
Authors: Sumit S. Shevtekar, Chandresh K. Maurya, Gourab Sil, Subasish Das
Submitted: 18 Aug 2026
Subjects: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Human-Computer Interaction (cs.HC)

arXiv: 2608.17823
Authors: Sumit S. Shevtekar, Chandresh K. Maurya, Gourab Sil, Subasish Das
Submitted: 18 Aug 2026
Subjects: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Human-Computer Interaction (cs.HC)


Summary

Summary

MotoSafety 是一种轻量级、高性能的边缘AI架构,专为评估两轮机动车(PTW)的碰撞风险而设计,特别针对诸如时间压力(TP)等认知应激因素的影响进行了优化。通过利用新颖的“学习型时间重要性”原则,该模型在分类和预测任务上均实现了 SOTA(业界领先)的性能,同时保持了足以部署在低成本边缘硬件上的高效率。

MotoSafety is a lightweight, high-performance Edge-AI architecture designed to assess collision risk for powered two-wheelers (PTWs), specifically addressing the impact of cognitive stressors like Time Pressure (TP). By utilizing a novel "Learned Temporal Importance" principle, the model achieves state-of-the-art performance in both classification and forecasting while remaining efficient enough for deployment on low-cost edge hardware.


Key Highlights

Key Highlights

1. Dataset & Methodology

1. Dataset & Methodology

  • 大规模数据: 本研究引入了一个包含超过 129,000 个标记多元时间序列序列的数据集,该数据集源自 153 次模拟器骑行(51 名参与者)。
  • 特征集: 捕获了 64 个不同的特征,包括车辆动力学、控制输入、邻近度和行为违规情况。
  • 归纳偏置: 将真实的 TP 作为归纳偏置引入,使分类准确率从 94.09% 提升至 94.97%。
  • Large-Scale Data: The study introduces a dataset comprising over 129,000 labeled multivariate time-series sequences derived from 153 simulator rides (51 participants).
  • Feature Set: Captures 64 distinct features, including vehicle dynamics, control inputs, proximity, and behavioral violations.
  • Inductive Bias: Incorporating ground-truth TP as an inductive bias improved classification accuracy from 94.09% to 94.97%.

2. Performance Metrics

2. Performance Metrics

  • 分类性能: 准确率达到 94.97%,ROC AUC 达到 99.33%,优于包括 TimesNetLLM4TS 在内的十个基线模型。
  • 预测性能: 实现了 0.039 的 MSE 和 0.094 的 MAE,与 Time-LLMiTransformer 相比,误差降低了 4.4 倍。
  • 效率: 该模型仅拥有 115 万个参数,运行延迟仅为 0.135 毫秒,非常适合低成本 CPU 边缘设备部署。
  • Classification: 94.97% accuracy and 99.33% ROC AUC, outperforming ten baselines, including TimesNet and LLM4TS.
  • Forecasting: Achieved 0.039 MSE and 0.094 MAE, representing a 4.4x reduction in error compared to Time-LLM and iTransformer.
  • Efficiency: The model features only 1.15M parameters and operates with a latency of just 0.135 ms, making it ideal for low-cost CPU edge deployment.

3. Practicality & Transferability

3. Practicality & Transferability

  • 即用型部署: 该模型仅使用 21 个 IMU+GPS 特征就保持了高准确率(93.91%),从而促进了真实场景中的落地实施。
  • 领域通用性: 除了 PTW 安全领域外,该架构在人类活动识别(97.66%)和临床领域(99.65%)中也表现出强大的迁移能力。
  • Deployment Ready: The model maintains high accuracy (93.91%) using only 21 IMU+GPS features, facilitating real-world implementation.
  • Domain Versatility: Beyond PTW safety, the architecture demonstrates strong transferability to human activity recognition (97.66%) and clinical domains (99.65%).

Access & Resources

Access & Resources


Metadata

Metadata