Generating High-Quality Vehicle Trajectories from Multiple Video Cameras

An advanced machine vision system is presented that automatically extracts vehicle trajectories over long freeway sections from video data. For each vehicle, its dimension and initial position are robustly given by the vehicle detection algorithm, and a separate tracking algorithm is used to extract the vehicle’s trajectory from the images of multiple cameras. The algorithms overcome the limitations of machine vision algorithms previously proposed that have difficulty to robustly obtain accurate vehicle positions in the presence of shadows and occlusions. Using the system, we generated a prototype dataset of over 4,700 individual vehicles, or 2.8M data points, which is the largest and most comprehensive dataset on vehicle trajectories ever produced.

Visit Federal Highway Administration's NGSIM Project for publicly available trajectory data and source code.

Input Images

Flow Diagram

Vehicle Detection

Model Fitting based on probabilistic line feature grouping		Truck Model: Input Lines and Example Detection Result:
Other Detection Results:

Vehicle Tracking

Tracking is based on the correlation matching of intensity pixels. Tracking parameters were tuned by systematic evaluation. For a complete set of trajectories, user-interface was developed to compensate mis-detection and tracking failure. User input is minimized due to efficient user-interface and good detection and tracking performance.

Publications

[Home]