ON FUSION-BASED APPROACHES FOR DETECTING, CLASSIFYING AND TRACKING OBJECTS FROM IMAGE AND VIDEO DATA

Date of Award

December 2018

Degree Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Electrical Engineering and Computer Science

Advisor(s)

Senem Velipasalar

Second Advisor

Cliff Davidson

Keywords

Object Detection, Object Tracking, Obstacle Classification, Obstacle Detection

Subject Categories

Engineering

Abstract

This thesis proposes and applies numerous fusion based approaches in image and video processing for understanding both objects and scenes. Fusion based approaches are those techniques that aggregate, ensemble, and combine decisions. These decisions are across numerous image and video based tasks. These image and video processing tasks are critical to leveraging image and video data in a wide range of applications, to include image search and retrieval, obstacle detection, and autonomous navigation. We demonstrate that the grouping of feature points is more accurately performed when the support for planar models is estimated based on the frequency counts of feature points preferences for a model. That is, a set of feature points prefer a model if the transformation of the feature points between two images is described by the model within an acceptable error rate. This approach represents a grouping supported by a voting of the feature points for these planar models. We similarly demonstrate for an object detection method how the voting of image edge pixels creates a more accurate scale estimate of the object. Next, we show that a selective agreement of multiple parts-based trackers is a technique that improves accuracies for the application of online tracking. We also show how a novel sensor setup and a deep-learning based approach achieve highly accurate obstacle detection and classification results. These results are further improved by smoothing and aggregating decisions over multiple frames. Finally, we discuss our work on creating an ensemble of learners by modeling the task of combining learners as a deep reinforcement learning task. Our results demonstrate across all of the proposed approaches, that the common theme of

information and decision aggregation consistently improves the outcomes across each application.

Access

SURFACE provides description only. Full text may be available to ProQuest subscribers. Please ask your Librarian for assistance.

This document is currently not available here.

Share

COinS