A Pipeline for Traffic Accident Dataset Development

P. 621–626.

Stepanyants V., Mantsa Andzhusheva, Romanov A.

Many traffic accidents happen on the roads every day and a lot of them are captured on traffic or dashboard cameras. This data could be used to train machine learning models to predict dangerous situations so that they can be prevented. For that, it should be organized into datasets. Nowadays a limited amount of traffic accident datasets is available and those are not as well annotated as, for example, driving datasets with no accidents, used for training automated vehicles. Following this, our paper presents a review of existing traffic accident datasets. The search was carried out to provide a list of video datasets relevant to the analysis of dangerous situations involving vehicles. For each dataset under consideration, a brief description of the process of data collection and annotation is presented, and the structure and format of videos are analyzed. In addition, the sources of the video, their amount, and the method of splitting the video into fragments are indicated. Where possible, software tools used for video processing are listed. Further, the paper explores existing solutions for dataset development and annotation. Based on the performed analysis, we propose a pipeline for traffic accident dataset development.

Keywords: segmentation, traffic accidents, detection, Dataset, video annotation

In book

2023 International Russian Smart Industry Conference (SmartIndustryCon), 27-31 March 2023

Sochi: IEEE, 2023.