Automatic Vehicle Detection and Tracking in Aerial Video

Abstract

In this report, we present an approach for the automatic detection and tracking of vehicles in aerial video. The proposed system is comprised of three stages namely; Region of Interest (ROI) Selection, Classification and Tracking. In the first stage we expedite detection by identifying image regions that are most likely to contain vehicles. To achieve this we use a fast corner detector combined with an efficient feature density estimation technique. In the classification stage the system uses shape encoding local features and robust machine learning techniques to perform efficient vehicle detection. In the final stage we propose a tracking algorithm, which employs particle filtering to exploit vehicle dynamics, and improve tracker-detector associations. We evaluate the proposed system by performing a number of experiments and use the results to show that the system is capable of successfully detecting and tracking a variety of differing vehicle types under varying rotation, sheering and blurring conditions.