Slot Cars: 3D Modelling for Improved Visual Traffic Analytics

Abstract

A major challenge in visual highway traffic analytics is to disaggregate individual vehicles from clusters formed in dense traffic conditions. Here we introduce a data driven 3D generative reasoning method to tackle this segmentation problem. The method is comprised of offline (learning) and online (inference) stages. In the offline stage, we fit a mixture model for the prior distribution of vehicle dimensions to labelled data. Given camera intrinsic parameters and height, we use a parallelism method to estimate highway lane structure and camera tilt to project 3D models to the image. In the online stage, foreground vehicle cluster segments are extracted using motion and background subtraction. For each segment, we use a data-driven MCMC method to estimate the vehicles configuration and dimensions that provide the most likely account of the observed foreground pixels. We evaluate the method on two highway datasets and demonstrate a substantial improvement on the state of the art.