Component-based car detection in street scene images

Abstract

Recent studies in object detection have shown that a component-based approach is more resilient to partial occlusions of objects, and more robust to natural pose variations, than the traditional global holistic approach. In this thesis, we consider the task of building a component-based detector in a more difficult domain: cars in natural images of street scenes. We demonstrate reasonable results for two different component-based systems, despite the large inherent variability of cars in these scenes. First, we present a car classification scheme based on learning similarities to features extracted by an interest operator. We then compare this system to traditional global approaches that use Support Vector Machines (SVMs). Finally, we present the design and implementation of a system to locate cars based on the detections of human-specified components.