Videodan yol ve trafik analizi

This thesis proposes two video-based traffic analysis systems, one for trafficmonitoring with fixed cameras, and one for driver warning applications with on-boardcameras looking outwards from the windshield. In the fixed-camera traffic monitoringsystem, Gaussian Mixture Model (GMM) based background subtraction is applied with anew adaptive bounding box size criteria to detect and track vehicles. An automaticallyextracted road mask is used to reduce the computational complexity. Furthermore, a newocclusion reasoning algorithm is proposed for robust tracking and counting of vehicles,where features such as size and width of the vehicles are used. Proposed system is testedunder different lighting and weather conditions, such as night and winter recordings. Inthe driver warning system with on-board camera, host vehicle localization with respect tolane marks and vehicle-to-vehicle distance calculation are addressed. A feature basedlane mark detection scheme with two step aggregation method is proposed. Edge featuresare used in this study, and aggregated into more meaningful structures by a new two-stepaggregation method. Tracking of the individual lane mark is handled by four Kalmanfilters for each of the lane mark corner. After analyzing the tracking results, two modesare defined for the host vehicle: in-lane and passing modes. Reliability of the proposedsystem is tested for host vehicle localization and vehicle-to-vehicle distance on a videosequence including both modes. Moreover, a new scene initialization procedure based onglobal motion estimation is used in this study. Experimental results show that theproposed algorithms perform well, and they are robust to environmental conditions.