|
Vitaladevuni Shiv Naga Prasad
|
||||
|
I am working in computer vision, and my research interests include human movement analysis, shape description and perceptual organization. Here is my resume. Edge Grouping for Contour Matching: A number of computer vision approaches rely on contour and gradient matching for detecting humans and pose-estimation. Edge clutter present in natural images results in erroneous matches as the detector can match with a subset of an object's edges and ignore the rest of them. We have proposed an approach for coupling edge grouping with contour matching. Experimental results indicate that this significantly improves pose-estimation, and reduces false human detections. Human Movement Analysis: This work addresses the problem of recognizing human reach and strike movements. These movements have highly variable targets - a subject can reach for something on the floor, above his/her head, etc. The challenge is to model the factors common to reach movements that are independent of the movement's target location. Psychological studies indicate that human reaches and strikes are ballistic in nature. They have simple trajectories, and the whole body of the subject moves in synchrony with the hand. Explicit consideration of the ballistic dynamics, results in robust recognition and generalization over movement styles and target locations . This approach has been applied for analyzing human motion capture data and single-camera videos. Demo videos. MATLAB software for temporal segmentation (ZIP) Bilateral and Rotational Symmetry Detection: We have developed algorithms for efficiently detecting bilateral and rotational symmetries in natural images using differential geometry. Computer Generated Chart Analysis: We have developed a system for classifying computer generated charts into categories e.g., bar-charts, curve-plots. Such a categorization would have applications in document retrieval, semantic analysis of chart images, etc. This task is challenging due to the large intra-class variation - both in terms of structure and style. We employ perceptual grouping techniques to extract salient curves and regions present in the images, and characterize their shape. Chart images are classified based on their similarity with example images of each category, measured by the overlap in the distributions of the features. We have tested the approach with more than 650 images collected from the Internet. Click here to view the details.
Publications:
Reports:
Downloads:
|
||||