Previous Projects

High-Quality Capture of Documents on a Cluttered Tabletop with a 4K Video Camera

We present a novel system for detecting and capturing paper documents on a tabletop using a 4K video camera mounted overhead on pan-tilt servos. Our automated system first finds paper documents on a cluttered tabletop based on a text probability map, and then takes a sequence of high-resolution frames of the located document to reconstruct a high-quality, fronto-parallel document page image. The quality of the resulting images enables OCR processing on the whole page. In a preliminary evaluation on 10 document pages, the system achieved 98% accuracy with the open-source Tesseract OCR engine.

Visual Saliency in Noisy Images

The human visual system possesses the remarkable ability to pick out salient objects in images, and it keeps this ability despite noise, poor weather, and other impediments to perfect vision. Noise, however, can significantly degrade the accuracy of automated saliency detection. Existing computational saliency models generally assume that the given image is clean, and a fundamental, explicit treatment of saliency in noisy images was missing from the literature.

We propose a statistically sound method for estimating saliency based on a non-parametric regression framework, and analyze how state-of-the-art models respond to noisy visual stimuli. Saliency at a pixel is modeled as a data-dependent weighted average of dissimilarities between a center patch around that pixel and other patches. To improve both fixation prediction and stability to noise, we extend the local analysis window to the entire image and to multiple scaled copies of it. Our method consistently outperforms six other state-of-the-art models in both noise-free and noisy cases.

Dependable Statistical Fusion for 3D Object Recognition and Pose Estimation

A probabilistic framework for recognizing 3D objects and estimating their pose in real environments. The approach features (1) automatic selection and collection of an optimal set of evidence based on in-situ monitoring of environmental variations, (2) multiple interpretations, as particles representing possible object poses in 3D space, with probabilities assigned by matching the object model against the evidence, and (3) fusion of interpretations over time with additional evidence from a sequence of images.

Using Structured Light for Efficient Depth Edge Detection

A novel approach that accurately detects depth edges while ignoring cluttered inner texture edges. We strategically project structured light and exploit the distortion of the light pattern along depth discontinuities to reliably detect depth edges. Depending on the distance from the camera or projector, this distortion may not occur or may be too small to detect, so we present methods that guarantee it occurs for a continuous range of object locations. Experiments show that the method accurately detects depth edges of human hands and bodies as well as general objects.

Machine Vision Inspection Systems

Vision-based profile monitoring of plastic tubes

Six calibrated camera-laser units monitor deformation of a plastic tube. A laser line projected onto the surface appears distorted by its 3D shape, revealing a partial surface profile from each unit; the profiles are combined to recover the whole tube profile.

Bullet defect detection

Dents, dirt, and smears on bullet casings are detected by projecting stripe light patterns onto the surface and observing distortions in the patterns.