Car Detection by Classification of Image Segments
Halpenny, Robert Morgan
COMPUTER SCIENCE CAR DETECTION BY CLASSIFICATION OF IMAGE SEGMENTS Robert Morgan Halpenny Thesis under the direction of Professor Xenofon D. Koutsoukos Object detection is one of the most important unsolved problems in computer vision. Even recognizing a single class of object (such as a car), is complicated by object variation, changes in lighting and perspective, and object occlusion. In this approach, we attempt to detect cars in images of city streets by classifying image segments based on SIFT keypoints. SIFT keypoints provide image features that are strongly resistant to changes in perspective and lighting. We learn the most predictive keypoints by applying Bayesian inductive learning via the HITON-PC algorithm. Each segment is classified by a Support-Vector Machine (SVM) using the keypoints in and adjacent to the segment. Classifying segments allows us to greatly reduce the scope of classification efforts while retaining high cohesion among intra-segment keypoints, yielding a fast and highly predictive detection algorithm.