TU-DAT: A Computer Vision Dataset on Road Traffic Anomalies

This paper introduces a novel, freely downloadable computer vision dataset for analyzing traffic accidents using roadside cameras. Our objective is to address the lack of public data for research concerning the automatic detection and prediction of road anomalies and accidents to enhance traffic safety. The Temple University Data on Anomalous Traffic (TU-DAT) dataset comprises various accident videos from news reporting and documentary websites. To guarantee the applicability of our method to roadside edge devices, we exclusively utilize footage and images from traffic CCTV cameras. We have collected approximately 210 videos, ranging from 24 to 30 FPS, depicting road accidents, comprising 17,255 accident keyframes and 505,245 standard frames. Analysis of the TU-DAT dataset revealed a significant finding. Due to the challenges in acquiring real-world traffic videos to analyze aggressive driving, we used a game simulator to produce road traffic video data that emulates aggressive driving behaviors, including speeding, tailgating, weaving through traffic, and disregarding red lights. We collected around 40 videos of positive instances and 25 videos of negative cases. We have already used this dataset in several contexts where we integrate deep learning with explicit spatiotemporal logic reasoning and demonstrate substantial outperformance over pure deep-learning methods in accuracy and running time. We hope it will be used innovatively for computer vision research.