Scholarlink

<p>This work proposed an Adaptive Temporal-Visual Hybrid Network (ATVHN) suited for precise conflict detection and threat estimation in harsh road scenarios for traffic safety. Conventional traffic systems are sometimes unable to trace the subtle relations between cars and pedestrians, particularly under difficult environments like poor light, occlusion, and anarchic motion. The envisioned framework combines a pre-trained YOLOv8 model with multi-scale video analysis and adds two new attention modules: Adaptive Visual-Attention Embedding Network (AVAENet) and Hybrid Temporal Dynamics Attention Network (HyTeDANet). These modules extract key spatial and temporal features from HWID12 video and time-series data, enriched through preprocessing and fusion techniques. Experimental performance indicates that ATVHN performs better than current models, such as Bi-LSTM and RNN, with high accuracy (0.98), precision (0.97), and sensitivity (0.97). Strong performance notwithstanding, the model is limited in effectiveness by sensitivity to high-quality video input and computational burden. Possible applications include integration within intelligent traffic systems for real-time surveillance, risk prediction alerts, and adaptive city traffic planning.</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

Advanced Traffic Conflict Detection and Risk Assessment Using Multi-Scale Video Analysis: A YOLOv8 Modified and Attention-Enhanced Safety Metrics Evaluation

  • 期刊论文

  • 发表日期：

    2025年6月18日

查看全文
  • Avantika Singh,
  • Sachin Dass

摘要

This work proposed an Adaptive Temporal-Visual Hybrid Network (ATVHN) suited for precise conflict detection and threat estimation in harsh road scenarios for traffic safety. Conventional traffic systems are sometimes unable to trace the subtle relations between cars and pedestrians, particularly under difficult environments like poor light, occlusion, and anarchic motion. The envisioned framework combines a pre-trained YOLOv8 model with multi-scale video analysis and adds two new attention modules: Adaptive Visual-Attention Embedding Network (AVAENet) and Hybrid Temporal Dynamics Attention Network (HyTeDANet). These modules extract key spatial and temporal features from HWID12 video and time-series data, enriched through preprocessing and fusion techniques. Experimental performance indicates that ATVHN performs better than current models, such as Bi-LSTM and RNN, with high accuracy (0.98), precision (0.97), and sensitivity (0.97). Strong performance notwithstanding, the model is limited in effectiveness by sensitivity to high-quality video input and computational burden. Possible applications include integration within intelligent traffic systems for real-time surveillance, risk prediction alerts, and adaptive city traffic planning.