Scholarlink

<p>This paper presents a precipitation-aware sensor ecosystem modelling framework designed to enhance autonomous vehicle (AV) navigation under adverse weather. By integrating spatiotemporal radar-based precipitation nowcasting, deep neural network-based sensor degradation modelling, and adaptive probabilistic sensor fusion, the system dynamically adjusts LiDAR, RADAR, and camera inputs according to precipitation severity and sensor reliability. The proposed framework is evaluated against a fixed-weight baseline detector with uniform multi-sensor fusion under identical training and hardware conditions. On 4.5&#xa0;TB of multi-modal sensor data collected over 320&#xa0;driving hours across five precipitation conditions, the framework achieved a 31.2% increase in detection precision, a 27.8% reduction in false positives, and a reduction in mean perception latency from 51 to 43&#xa0;ms. Radar-based rainfall estimation yielded a mean absolute error of 0.42&#xa0;mm/h with <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(R^{2}=0.91\)</EquationSource> </InlineEquation>. Key metrics are reported as mean ± standard deviation over repeated evaluation runs and are further summarized with class-wise Average Precision and statistical variability in the Results section. The dataset used in this study is a custom multi-modal dataset collected during the project and is available from the authors upon reasonable request.</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

Precipitation-aware sensor ecosystem modelling for performance-driven autonomous vehicle navigation

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月19日

查看全文
  • Shruti Kalra,
  • Ruby Beniwal,
  • Narendra Singh Beniwal,
  • Maninder Singh

摘要

This paper presents a precipitation-aware sensor ecosystem modelling framework designed to enhance autonomous vehicle (AV) navigation under adverse weather. By integrating spatiotemporal radar-based precipitation nowcasting, deep neural network-based sensor degradation modelling, and adaptive probabilistic sensor fusion, the system dynamically adjusts LiDAR, RADAR, and camera inputs according to precipitation severity and sensor reliability. The proposed framework is evaluated against a fixed-weight baseline detector with uniform multi-sensor fusion under identical training and hardware conditions. On 4.5 TB of multi-modal sensor data collected over 320 driving hours across five precipitation conditions, the framework achieved a 31.2% increase in detection precision, a 27.8% reduction in false positives, and a reduction in mean perception latency from 51 to 43 ms. Radar-based rainfall estimation yielded a mean absolute error of 0.42 mm/h with \(R^{2}=0.91\) . Key metrics are reported as mean ± standard deviation over repeated evaluation runs and are further summarized with class-wise Average Precision and statistical variability in the Results section. The dataset used in this study is a custom multi-modal dataset collected during the project and is available from the authors upon reasonable request.