A description is given of the system architecture of an autonomous vehicle and its real-time adaptive vision system for road-following. The vehicle is a 10-ton armored personnel carrier modified for robotic control. A color transformation that best discriminates road and nonroad regions is derived from labeled data samples. A maximum-likelihood pixel classification technique is then used to classify pixels in the transformed color image. The vision system adapts itself to road changes in two ways; color transformation parameters are updated infrequently to accommodate significant road color changes, and classifier parameters are updated every processing cycle to deal with gradual color and intensity changes. To reduce unnecessary computation, only the most likely road region in the segmented image is selected, and a polygonal representation of the detected road region boundary is transformed from the image coordinate system to the local vehicle coordinate system based on a flat-earth assumption.>

The major difficulty in stereo vision is the correspondence problem that requires matching features in two stereo images. In this paper, we describe a constraint-based stereo matching technique using local geometric constraints among edge segments to limit the search space and to resolve matching ambiguity. Edge segment is used as the primitive image feature for stereo matching. Standard epipolar constraint and individual edge properties are used to determine possible initial matches between edge segments in the left and right images. Local edge geometric relations such as continuity, junction structure, and edge neighborhood relations are then used as constraints to guide the stereo matching process. The result is a locally consistent set of edge segment correspondences between stereo images. These locally consistent matches then generate higher-level hypotheses on extended edge segments and junctions. These hypotheses form more global contexts and allow local supports to correct incorrect local matches to achieve global consistency.

This paper describes an algorithm that decomposes the free space into nonoverlapping geometric-shaped primitives suitable for path planning. Given a set of polygonal obstacles in space, the algorithm first decomposes concave obstacles into connecting convex obstacles in order to have a uniform obstacle representation, which facilitates later processing. The neighborhood relations among these Convex obstacles are identified and then used to locate critical "channels" and "passage regions" in the free space, and to localize the influence of obstacles on free space description. The free space is then decomposed into nonoverlapping geometric-shaped primitives where channels connect passage regions. The channels are similar to the generalized cones presented in Brooks [1]. The passage regions are represented as convex polygons. Based on this mixed representation of free space, the path planning algorithm can plan path trajectories inside the channels and passage regions. The perimeter of the channels and passage regions associated with the planned path provides the boundaries of the path that the mobile robot must stay within during path execution.

In this paper, we consider the restoration of images with signal-dependent noise. The filter is noise smoothing and adapts to local changes in image statistics based on a nonstationary mean, nonstationary variance (NMNV) image model. For images degraded by a class of uncorrelated, signal-dependent noise without blur, the adaptive noise smoothing filter becomes a point processor and is similar to Lee's local statistics algorithm [16]. The filter is able to adapt itself to the nonstationary local image statistics in the presence of different types of signal-dependent noise. For multiplicative noise, the adaptive noise smoothing filter is a systematic derivation of Lee's algorithm with some extensions that allow different estimators for the local image variance. The advantage of the derivation is its easy extension to deal with various types of signal-dependent noise. Film-grain and Poisson signal-dependent restoration problems are also considered as examples. All the nonstationary image statistical parameters needed for the filter can be estimated from the noisy image and no a priori information about the original image is required.

A nonstationary 2-D recursive image restoration filter that uses a nonstationary mean, nonstationary variance (NMNV) image model and minimizes the local mean square error is developed. The 2-D recursive filter adapts itself to the local image statistics and is able to do space-variant processing. The NMNV image model has a simple dynamic representation which simplifies the filter structure considerably. However, the optimal recursive filter still requires extensive computation. A suboptimal approach that uses a reduced update concept is proposed to reduce the computational efforts. With some modifications, this nonstationary 2-D recursive filter is extended to a class of uncorrelated, signal-dependent noise such as multiplicative noise and Poisson noise. The explicit filter structures and simulation results for images degraded by these signal-dependent noises are presented.

Speckle noise exists in all types of coherent imagery such as synthetic aperture radar, acoustic imagery and laser illuminated imagery. Speckle can be reduced by averaging over several uncorrelated speckle images of the same object when these are available. In this paper, we attempt to reduce speckle noise from a single speckle image by using adaptive digital image restoration techniques. Many speckle noise reduction algorithms assume speckle noise is multiplicative. We model the speckle according the exact physical process of coherent image formation. Thus, the model includes signal-dependent effects and accurately represents the statistical properties of speckle. A linear minimum mean-square error filter is derived based on our speckle model and a nonstationary image model. The filter responds adaptively to the signal-dependent speckle noise and the nonstationary mean and variance of the original image. The necessary parameters are estimated from the noisy image. The 2-D recursive implementation of this filter is developed as a fast computation algorithm.