A Rubust System for Automated Video-Based Vehicle Recognition

We propose an end-to-end video-based system for online class recognition of moving vehicles observed by a passive stationary camera. SAVOR (or System for Automated Video-based Object Recognition) continuously detects, tracks, and classifies vehicles within the camera's field of view in real time, additionally providing a live operator display and archiving data and results for later analysis. To our knowledge, SAVOR will represent the first continuously-operating passive vehicle recognition system in existence. The system augments a proven image-based Predict-Extract-Match-Search (PEMS) framework, employing pre-defined shape and appearance templates and exploiting the spatio-temporal coherence inherent in motion imagery to extract and correlate rich two- and three-dimensional feature sets. Calibration of geometric and photometric environmental attributes allows robust operation and consistent reasoning over a constrained metric parameter space, while a sophisticated appearance prediction engine able to account for such phenomena as reflectance and cast shadows additionally enables highly accurate template-to-image matching. SAVOR is designed to approach a vehicle type classification rate of 95% or greater on ordinary daytime traffic. Real-time operation allows statistically meaningful performance analyses over long durations and in temporally-varying environmental conditions. We will develop simple graphical interfaces for data analysis and truthing, and assess the system using isolated offline test cases for repeatable algorithmic evaluation as well as the live data stream for long-term analysis.