Understanding Corner Cases for Vehicle Detection and Tracking


At a glance

Vehicle detection and tracking has been an important problem in computer vision. Prominent applications include autonomous driving, remote sensing, and traffic surveillance. While vehicle detection and tracking have mostly been solved for most of daily scenarios, it still remains an open problem in corner cases like low-light condition, foggy or rainy weather, optical flares, occlusion, and nested geometry. In response to this, we propose to build a video dataset, using the I-24 MOTION testbed. The testbed is an array of traffic surveillance camera poles in Nashville, TN. The poles operate 24/7, span over 4 miles of highway, and contain 294 ultra-high definition cameras. Interesting clips of video footage will be selected and annotated by human annotators from the camera live feed. By sampling those corner cases, we hope the dataset could facilitate evaluation and improvement of the state-of-the-art models for vehicle detection and tracking.

Principal investigatorsresearchersthemes
Alexandre Bayen

Fangyu Wu

Jonathan Lee

Vehicle detection, Vehicle tracking, Computer vision dataset