CSE 252C: Selected Topics in Vision and Learning

Project Reports

December 2005

Robin Hewitt: Face Recognition for Mobile Robots

Abstract: A method for detecting and recognizing faces from one input image was implemented. This method uses a cascaded search with haarlike features to quickly eliminate image regions that do not match the input model. The remaining regions are evaluated at three scales using image-gradient features. Closeness to the input model is evaluated as the sum of local deformations to bring the search region and the input model into correspondence. The resulting method is computationally efficient and was able to recognize the original face under moderate changes in lighting, pose, and expression with a true-positive rate of 85% when the threshold for per-frame false-positive rate was adjusted to be 15%.

Anup Doshi: People Counting and Tracking for Surveillance

Abstract: The computer vision community has expended a great amount of effort in recent years towards the goal of tracking people in videos. Much more recently, algorithms have been developed to track multiple people in videos robustly and in real-time. The goal of this project is to implement a system based on one of those algorithms, in order to count and track the people in a database of surveillance footage. Due to several constraints and performance issues, however, a more straightforward algorithm based on background subtraction is implemented and shows acceptable performance levels. Further improvements are considered to improve the performance, including implementations of algorithms such as BraMBLe.

Ben Laxton: A Parts-based Person Detector

Abstract: This work presents a parts-based person detection framework. The individual body segments- legs, torsos, and heads - are detected by Adaboost classifiers, which have proven useful other classification tasks. The body segment candidates are combined into kinematically plausible configurations and each is assigned a score. The body configurations are calculated using an efficient dynamic programming implementation for pictorial structures. The tracker is run on several video sequences that exhibit camera motion and significant differences in lighting and environment. This method can be made to run in realtime and can be easily extended to incorporate more complex models of human kinematics.

David Torres: More Local Structure Information for Better Make-Model Recognition

Abstract: An object classification technique is proposed to solve a vehicle make and model recognition task. Edges of the back end of vehicles are extracted from images. These edges are processed into line segments which contain more local structure information than interest point based characterization can encode. Object matching is performed by comparing the sets of line segments by a Hausdorff distance. The method is tested on a database of vehicle images.

Most recently updated on Nov. 29, 2005 by Serge Belongie.