Low-level computer vision applications to surveillance and robotics.

摘要

Computer Vision is an important component of environment sensing systems for applications such as robotics and surveillance. Images and video sequences contain information about the scene at several different levels. Accordingly, computer vision algorithms are traditionally classified into a hierarchy of levels, from low-level (e.g., feature extraction and matching) to high-level (e.g., object recognition).; This dissertation is developing a number of low level vision algorithms with applications to surveillance and robotics. The first part of this dissertation addresses automatic feature matching for camera pair calibration. An algorithm was proposed that uses a novel topological constraint, named "cross epipolar ordering constraint" (CEO constraint). No knowledge of camera orientation or position is needed for this constraint. It effectively reduces the search space for feature matching by ruling out matches that are proven to be unrealizable. An wide baseline feature matching algorithm is proposed using this constraint. This algorithm uses coarse-to-fine search, and reduces the search space further more.; The second part of this dissertation describes work on the analysis of stereo data for the automatic detection of "curbs" and "steps". Our real-time algorithm is able to determine the 3-D location of a curb, which uses robust techniques to combine range and brightness information. This algorithm was developed as part of a DARPA project for urban robotics, and is currently being ported to a semi-autonomous wheelchair control system.; The third part of this dissertation focus on a fast motion detection algorithm on a power-awared visual sensor network developed at UC Santa Cruz, Meerkats. This algorithm can detect moving blobs in two consecutive images and calculate the velocity of the motion using optical flow constraint. It is based on the rank analysis of the structure matrix, which is built up from local spatio-temporal gradients of a image sequence. Then a multiscale belief propagation scheme is also proposed in the dissertation. The multiscale belief propagation can deal with different sized moving blobs and large motions, which have been problems in optical flow calculation. This algorithm is designed specifically for Meerkats networks, so we give a detailed power consumption analysis of our algorithm on Meerkats.