Multi-perspective vehicle detection and tracking: Challenges, dataset, and metrics

Abstract

The research community has shown significant improvements in both vision-based detection and tracking of vehicles, working towards a high level understanding of on-road maneuvers. Behaviors of surrounding vehicles in a highway environment is found as an interesting starting point, of why this dataset is introduced along with its challenges and evaluation metrics. A vision-based multi-perspective dataset is presented, containing a full panoramic view from a moving platform driving on U.S. highways capturing 2704×1440 resolution images at 12 frames per second. The dataset serves multiple purposes to be used as traditional detection and tracking, together with tracking of vehicles across perspectives. Each of the four perspectives have been annotated, resulting in more than 4000 bounding boxes in order to evaluate and compare novel methods.