情報処理学会 第79回全国大会講演要旨

With the popularization of automobiles, the miss caused by drivers may lead to severe traffic accidents. In fact, an ability of danger detection during driving calls a high attention. As there is still no high qualified open dataset for driving danger detection, we aimed at producing a dataset of that. We have collected over 1600 video data from open source and annotated them with several danger information, including time of risk appearance and accident occurrence, danger category, direction and so on. Also, we tried existing methods, such as Max-Margin Early Detection, on automobile part of our original dataset to make sure of its quality.