Map-Aided Vehicle Tracking Used in Behaviroural Studies for Intersection Driver Assistance Applications

Abstract

This thesis addresses the vehicle tracking problem in video recordings in intersections for driver behavior studies with the assistance of digital maps. For object detection and tracking, both the own-vehicle and other vehicle motion parameters like position, speed, heading and acceleration are needed. This information is measured using sensors installed on a test vehicle. These include GPS, gyros, video cameras, and a digital map. The digital map is used as a sensor for improving vehicle tracking performance. Since, all sensors contain errors and uncertainties, error models are computed for some certain sensors. To compare different tracking strategies, performance measures are introduced. By using the manual object selection (from a Graphical User Interface) in a few video frames, at least in the beginning and end, an estimation algorithm computes this vehicle's position along the road. This map estimation technique was enhanced by adding an image processing system, which looks for the previously selected object in the next frame. In order to include the other vehicle's motion and sensors error model a particle filter based tracking algorithm is designed. This method increases the tracking performance in the case that only a few manual markings have been given. It is shown in this study that with only a few manual markings, then the tracking/fusion algorithm will increase the performance by exploiting visual features from images. On the other hand, when there are many manual markings available, this simple image analysis does not extract more information than the map estimator.