Environmental exploration via computer vision techniques.
详细信息
文摘
The first part of this dissertation presents a novel environmental exploration tool (virtual white cane) for the visually impaired. The virtual white cane is a range sensing device composed of a laser pointer and a miniature camera in an active triangulation structure. The device has the features of self-calibration and light condition adaptation. Users can use the device in real time for sensing the environment, measuring distances, and detecting surface discontinuities, such as curbs, walls and drop-offs. This is achieved by analyzing the range data collected while the user is periodically pivoting the device around. A number of models including various planar surface models and switching schemes are proposed and implemented. The planar surface model uses an Extended Kalman Filter (EKF) to track the surface. This filter is significantly robust and efficient in detecting jumps between disjointed surfaces, and also the surfaces perpendicular to each other. The scheme using Jump-Markov process and particle filtering with an optimal proposal distribution is for tracking switches between models. With the help of an inclinometer, the inclinations of the tool with respect to a reference direction is obtained which enables tracking the surface slant. A variety of experimental results prove the robustness of the tracking system in real world conditions.;The second part of this dissertation introduces a general framework for classification algorithm. The algorithm includes two phases: learning and testing. The learning phase takes some fundamental visual features (such as color, normalized color, texture) from segmented images for training the classifiers. The classifiers are a number of multivariate Gaussian models which describe the distributions of image segments. A merge procedure is performed between any two of these models if the Bhattacharya distance between these two distributions is smaller than a preset threshold. The testing phase measures distance between the candidate segments/regions/pixels in the test images and the learnt models, and classifies them into different categories of interest. Two derivatives with a few specific add-ons of this algorithm are applied to different ground robot navigation tasks in the LAGR project. The image data for the off-line experiments was taken in the real test site of each task. Both the off-line and real-time tests showed that the classification algorithm worked quite well in learning the outdoor and natural environmental features, then guiding the robots moving through the similar terrain until they reach the preset goals.