(click here for my personal webpages)
Howard Schultz, Ph.D.
Senior Research Scientist, Computer Vision Laboratory
University of Massachusetts
Computer Science Department
140 Governors Drive
Amherst, MA 01003-4610
office phone: (413) 545-3482
secretary: (413) 545-2746
fax: (413) 545-1249
I received my BS and MS degrees in Physics from UCLA in 1972 and 1974, and my Ph.D. in Physical Oceanography from the University of Michigan, Ann Arbor in 1982. I am a member of the IEEE, ASPRS and AGU. My research interests focus on quantitative methods for image understanding and remote sensing, including generating complex, three-dimensional terrain and site models from aerial images.
Environmental monitoring. The growing importance and visibility of global environmental studies has increased the demand for systems capable of collecting large amounts of high-resolution, content-rich aerial image and feature data. These systems often must be able to operate under difficult conditions in remote, uninhabited regions. Furthermore, the successful modeling and monitoring of complex biological systems requires portable, low-cost equipment that can be operated with a minimum amount of training. Although satellite images are often available, they typically have insufficient resolution for many applications, their coverage may be limited, they are often obstructed by clouds, and researchers cannot control the timing of the overflight. Conventional aerial imaging allows more flexibility in the timing, but these systems are very expensive and require intensive training. The primary objective of our Environmental Monitoring Program is to build a low-cost, user friendly, robust digital imaging system consisting of a light-weight instrument package and suite of robust processing algorithms that will enable environmental scientists to produce high-quality 3D terrain models just about anywhere in the world at extremely low cost.
Over the past several years, the University of Massachusetts Computer Vision Laboratory (CVL) has been engaged in several collaborations with environmental monitoring and ecosystem modeling research programs at other universities, governmental agencies, nonprofit institutions, and commercial organizations. Our goal is to enhance the capabilities of our research partners by applying state-of-the-art digital imaging techniques to their research domains. Our initial collaborations focused on providing tools for determining ground cover classification from multiple image sources that spanned a wide range of resolutions. To this end, we developed techniques for producing geographically registered mosaics from analog and digital video camcorders. These mosaics were then used as input to traditional supervised classification schemes, as well as a new interactive classification paradigm developed at UMass. The tremendous success of these collaborations has created a demand for more sophisticated techniques and data products. Many of our environmental research and management partners require the ability to rapidly generate three-dimensional, multi-spectral geographically registered terrain models with a high spatial and spectral accuracy. To meet these requirements, we are building a purely digital image based system for generating geographically registered 3D terrain models in the form of seamless ortho-rectified mosaics.
3D Terrain reconstruction. Many methods for modeling terrain are limited because they require manual intervention and highly calibrated, down-looking cameras. However, many environmental monitoring and defense applications require automatic generation of 3D elevation models and feature maps from images taken by a variety imaging devices with widely varying viewing angles. To address these limitations, I developed the Terrest Reconstruction System (TERREST) which automatically generates digital elevation maps (DEM) from two or more views of a terrain (Schultz, 1995, 1996, 1999). The system has been shown to work well the viewing angles become oblique (incidence angles greater than 45 degrees), and/or when the camera separation becomes large (base-to-height ratio exceeds 1.5). Terrest was the only university software tool selected by National Imagery and Mapping Administration (NIMA) as part of the Path Finder 2000 competition.
Assimilating existing 3D information. High resolution models often are desired for areas where low resolution terrain models exist, such as 30 meter USGS digital terrain models and radar range images. Rather than discard the low resolution models, which include accurate information, it would be advantageous to refine an existing model. I developed an approach that uses the existing model to initialize the calculation of the new high resolution model. This technique was shown to improve significantly the efficiency and robustness of the terrain modeling process (Schultz et al., 1997a, b).
Using 3D texture to improve classifier performance. Most pixel classification algorithms rely on 2D image texture information from a single image. When multiple images exist, however, additional information is available about the 3D structure of the terrain. Our approach (Wang, Stolle, and Schultz et al., 1997) defines a set of texture measures computed during image matching. These measure include the match score and the width of the similarity function. These 3D textures characterized the microstructure of the terrain. Experiments demonstrated that adding these 3D texture features to the standard set significantly improved the ability of a classification scheme to discriminate between different classes of ground cover.