CADP

我们提出了一个用于交通事故分析的新数据集。我们的目标是解决道路交通安全自动时空注释研究缺乏公共数据的问题。我们的车祸检测和预测~（CADP）数据集由从 YouTube 收集的 1,416 个视频片段组成，其中 205 个视频片段具有完整的时空注释。通过对 CADP 数据集的分析，我们观察到由于对象大小和场景的复杂性，我们数据集中行人类别的对象检测显着下降。为此，我们建议将增强上下文挖掘 (ACM) 集成到 Faster R-CNN 检测器中，以补充小行人检测的准确性。我们的实验表明，CM 的目标检测准确率显着提高 +8.51%，ACM 提高 +6.20%。对于人~（行人）类别，我们观察到显着的改进：与 Faster R-CNN 相比，CM 为~+46.45\%，ACM 为 +45.22\%。最后，我们使用改进的 Faster R-CNN 和 Accident LSTM 架构在我们的数据集中展示了事故预测的性能。就事故发生时间而言，我们平均达到了 1.684 秒，最高平均精度为 47.25%。我们希望我们的数据集可以作为新研究方向的起点，并在未来几年逐步增长

We present a novel dataset for traffic accident analysis, aiming to address the shortage of public datasets for research on automated spatio-temporal annotation of road traffic safety. Our Crash Detection and Prediction (CADP) dataset consists of 1,416 video clips collected from YouTube, among which 205 clips have complete spatio-temporal annotations. Through analysis of the CADP dataset, we observed that object detection performance for the pedestrian category in our dataset experiences a significant decline due to small object sizes and complex scene contexts. To this end, we propose integrating Augmented Context Mining (ACM) into the Faster R-CNN detector to improve the detection accuracy of small pedestrians. Our experimental results demonstrate that the overall object detection accuracy of CM improves by 8.51% significantly, while that of ACM improves by 6.20%. For the pedestrian category, we observe notable improvements: compared with the baseline Faster R-CNN, CM achieves a 46.45% performance boost and ACM achieves a 45.22% performance boost. Finally, we demonstrate the accident prediction performance on our dataset using the improved Faster R-CNN and Accident LSTM architecture. In terms of the lead time before accident occurrence, we achieve an average lead time of 1.684 seconds, with a peak mean average precision of 47.25%. We hope that our dataset can serve as a starting point for new research directions and will continue to grow incrementally in the coming years.

AI搜集汇总

背景与挑战

背景概述

CADP是一个用于交通事故分析的数据集，包含1,416个从YouTube收集的视频片段，其中205个带有完整的时空注释，旨在解决道路交通安全自动时空注释研究数据不足的问题。该数据集专注于行人检测和事故预测，通过集成增强上下文挖掘技术显著提升了行人检测准确性，并利用改进的Faster R-CNN和Accident LSTM架构实现了有效的事故预测性能。