Robust estimation of depth perception and image segmentation of monocular image sequences with known camera translation

• Main Author: Ilao, Joel P.
• Format: text
• Language: English
• Published: Animo Repository 2008
• Subjects: Computer vision; Image segmentation; Computer Sciences
• Online Access: https://animorepository.dlsu.edu.ph/etd_masteral/3878; https://animorepository.dlsu.edu.ph/context/etd_masteral/article/10716/viewcontent/CDTG004775_P.pdf
• Institution: De La Salle University

Monocular vision techniques use information taken from a single moving camera in inferring the 3-D structure of a camera observers environment. Compared to polynocular vision techniques, monocular vision techniques require less hardware and information about the camera geometry, in order to estimate relative depth. However, monocular vision is more prone to image noise and is computationally expensive. This research proposes an algorithm for depth estimation for use in mobile robotic navigation. Depth estimation in real-world image sequences of a visual scene captured by a single moving camera using optic flow information still suffer from accuracy problems due imperfection in optic flow estimates. Since the Structure from Motion problem of Monocular Vision is regarded as non-linear, the initial optic flow estimate, hence, is further enhanced using a novel approach of applying Extended Kalman Filter formulation on the corresponding divergence fields. Raw optic flow estimates of consecutive frames in any given image sequence were computed using the pyramidal Lucas-Kanade algorithm. The resulting optic flow field is then used as basis for estimating the 3-D scene structure via construction of a depth/range map. The depth maps were constructed following a Camus formulation assuming monocular image sequences captured by a camera undergoing uniform forward translation. These depth maps were refined further by application of median filtering as a post-processing mechanism. Standard tests on synthetic and real-world images indicate that the Extended Kalman Filter has been effective in making the depth estimation process consistent, most especially if the optic flow estimates of the initial frame were made very close to the ideal (57.8% and 14.7% reduction in the standard deviation of divergence magnitude error values for the Kalman-filtered divergence data with and without ground truth values, respectively, over that which used only raw optic flow data). The system developed, however, cannot still effectively apply to real-world image data due to limiting assumptions on the observer motion type and imaged surface orientation relative to the camera observers focal axis, as well as lack of textural content and ambient lighting noise.