This was a project I worked on at the Smith-Kettlewell Eye Research Institute in the summer of 2009 with James Coughlan. I took their existing idea of quickly finding a color target with a cell phone camera and added some advanced model verification, multi-scale processing and used to the accelerometer to prioritize scanning the horizon at full resolution. It really works! But blind subjects usually still prefer using braille... The platform was a Nokia N95 smartphone, and the application could detect targets consistently from 3-4 meters away, and could read a simple barcode with about 8 bits of data from 1-3 meters away. Source code and publications may appear here or on the SKI lab site at some point in the future...

Multiple object, multiple non-overlapping camera tracking

We worked with SAIC to implement tracking algorithms for multiple objects across multiple cameras with and without overlapping regions.

Multi-scale Retinex with color restoration

An implementation of the retinex algorithm has been implemented for processing video sequences. Retinex is a general enhancement algorithm based on homomorphic filtering. It has the ability to compress dynamic range, remove illumination changes, and enhance sharpness and contrast.

Source code (ZIP) This was a graduate multimedia class project.

Right: Several examples of Retinex enhancement of video frames.

Crowd Estimation from surveillance footage.

Methodology: Scalar information is obtained from various image-processing techniques and used to collect training data. Various fitting techniques (linear weights, neural networks, etc.) are used to fit the scalar data to the actual crowd size. We then use this model to estimate the crowd size in new images.

Files: BMVC Paper (PDF), ICPR Paper (PDF).

HASTE: High Altitude Stratospheric Telemetry Explorer

Objective: Senior Design Project Class (CMPE 123A/B)

Methodology: The design and construction of a weather balloon equipped with various sensors, GPS and amateur radio equipment. The main goal of the project is the design of a differential thermocouple sensor which will be accurate enough to measure temperature differences across 1 meter with 0.001 degree centigrade accuracy. When sampled at 100 Hz, this measurement can be linked to the index of refraction structure parameter (C_n^2) of the air. Our experiment is to create a system of obtaining a profile of C_n^2 at a fraction of the current cost ($3k vs. $70k). While our balloon is currently missing in action in the vicinity of Fresno, we have succeeding in building a proof of concept.

Active Tracking

Objective: This was my project for CMPE 264 (Computer Vision and Image Analysis). The objective was to get a camera to follow people around the room.

Methodology: First an object is specified by the user (usually a persons face. A modified mean shift tracker then tracks the object around the image. When the object approaches the edge of the frame, a signal is sent to the camera to turn towards the object.

Files: Presentation Slides (PPT), Source Code (ZIP (updated 7-12-05)). Please note that I take no responsibility for anything this code is used for. It is unsupported and mostly undocumented. It contains code from OpenCV, an open source library sponsored by Intel. All the licensing restrictions that apply to OpenCV also apply to my code. (BSD style license)