Scholarlink

<p> <?tk 4?>This study presents a spatiotemporal traffic speed prediction framework using high-resolution unmanned aerial vehicles (UAV)-based trajectory data under heterogeneous, lane-free traffic conditions on Indian highways. Two 300&#xa0;m highway sections (one curved, one straight) were divided into six 50&#xa0;m regions across each lane and analyzed at one-minute intervals to capture lane-wise speed dynamics. A tailored spatiotemporal matrix transformation was developed to structure historical speed data from adjacent upstream regions and time steps into 3D arrays, enabling sequential learning. A comparative evaluation of five models, random forest (RF), convolutional neural network (CNN), recurrent neural network (RNN), long short-term memory (LSTM), along with the statistical autoregressive integrated moving average (ARIMA) was conducted. A detailed spatiotemporal window selection process, supported by variable importance analysis and error surface evaluation, established a 20&#xa0;min temporal and three-region spatial look-back as optimal. Performance was assessed using RMSE, MAE, and MAPE. Among the models, RF demonstrated superior accuracy, with MAE ranging from 2.45 to 3.95, RMSE from 2.99 to 3.65, and MAPE from 5.0 to 6.80% encompass both spatial and temporal predictions across the curved and straight highway sections. The LSTM model followed as the next best-performing technique. Model accuracy slightly declined for shoulder lanes due to higher speed variance caused by frequent vehicle entries/exits. Generalization was assessed on a third independent site, where RF maintained strong predictive accuracy. The proposed framework offers actionable value for real-time speed control, congestion mitigation, and proactive traffic safety in mixed-traffic environments where traditional sensor networks are sparse or ineffective.</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

Short-term traffic speed prediction models for highways using UAV-based trajectory data and artificial intelligence

  • 期刊论文

  • 发表日期：

    2025年9月9日

查看全文
  • Vineet Jain,
  • Ashish Dhamaniya

摘要

This study presents a spatiotemporal traffic speed prediction framework using high-resolution unmanned aerial vehicles (UAV)-based trajectory data under heterogeneous, lane-free traffic conditions on Indian highways. Two 300 m highway sections (one curved, one straight) were divided into six 50 m regions across each lane and analyzed at one-minute intervals to capture lane-wise speed dynamics. A tailored spatiotemporal matrix transformation was developed to structure historical speed data from adjacent upstream regions and time steps into 3D arrays, enabling sequential learning. A comparative evaluation of five models, random forest (RF), convolutional neural network (CNN), recurrent neural network (RNN), long short-term memory (LSTM), along with the statistical autoregressive integrated moving average (ARIMA) was conducted. A detailed spatiotemporal window selection process, supported by variable importance analysis and error surface evaluation, established a 20 min temporal and three-region spatial look-back as optimal. Performance was assessed using RMSE, MAE, and MAPE. Among the models, RF demonstrated superior accuracy, with MAE ranging from 2.45 to 3.95, RMSE from 2.99 to 3.65, and MAPE from 5.0 to 6.80% encompass both spatial and temporal predictions across the curved and straight highway sections. The LSTM model followed as the next best-performing technique. Model accuracy slightly declined for shoulder lanes due to higher speed variance caused by frequent vehicle entries/exits. Generalization was assessed on a third independent site, where RF maintained strong predictive accuracy. The proposed framework offers actionable value for real-time speed control, congestion mitigation, and proactive traffic safety in mixed-traffic environments where traditional sensor networks are sparse or ineffective.