Date of Award

11-6-2019

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Systems Engineering

First Advisor

Xian Liu

Abstract

This dissertation aims to develop object detection approaches for an indoor navigation aid called smart cane. The goal of the object detection approaches is to provide object-level information of the environment to assist the visually impaired people to walk freely in the indoor environment. Although many object detection approaches have been proposed, object detection is still a challenging problem due to a large number of object categories, viewpoint changes, occlusions, cluttered backgrounds, etc. To address these challenges, first, an indoor 3D dataset is collected by using the smart cane navigation aid. Second, indoor objects are broadly divided into two types: the waypoint object and the nonstructural object. A waypoint object is a structural object that plays an important role in the indoor way-finding process, thus helping the visually impaired to move from one place to another in the hallway. The nonstructural object can be used by the visually impaired to avoid obstacles and interact with the environment. Third, three object detection approaches are designed to effectively detect indoor objects based on the collected object dataset. The first object detection approach aims to detect the waypoint object by integrating intensity and depth information in a hierarchical structure. The hierarchical structure identifies geometric properties in a coarse-to-fine manner with the help of visual information. Then, these geometric properties are used to search for waypoint candidates. The second approach uses a convolutional neural network structure to learn object representations from multimodal data for detecting nonstructural objects. This method uses three inputs computed from intensity image and depth data to learn a useful object representation. The fusing of multimodal features helps to improve detection accuracy. The third approach is a salient object detection method that aims to find the most conspicuous object region in an input scene. This method uses tensor to represent high dimensional image features. Then the image feature tensor is decomposed into a low-rank tensor and a sparse tensor in a heuristic manner. Finally, the salient object can be effectively detected by assuming that the redundant background is represented by the low-rank tensor while salient object information is included in the sparse tensor.

Download

Share

COinS