DEPTH ESTIMATION FROM MONOCULAR CAMERA IN URBAN ENVIRONMENT

Abstract

Depth/distance estimation is a central problem in machine vision and robotics. It has been tackled mostly using stereovision which requires the use of two cameras with a known baseline distance simulating human vision. Other robotic systems use a single camera that has to be fixed and calibrated (focal length, tilting and panning angles). This limitation prohibits the free rotation of the camera which is needed in traffic monitoring systems and in aiding visually impaired people. In contrast, this thesis uses a single camera with no calibration except for the camera height (which is assumed to be the users height) allowing free motion, rotation and focal length change of the camera. The proposed technique estimates the distance (from the camera to the objects point of intersection with the ground level) by enhancing the vanishing point calculation procedure by introducing.two new ad­ justment parameters. The proposed technique yields accurate results when tested on images captured from an accurately measured road, and promising results when compared with human estimation of the distances on general images.