Object recognition in noisy natural images
Date
2017
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
University of Delaware
Abstract
Autonomous robotic systems can operate in an unsupervised manner over remote
or potentially dangerous domains. Object recognition is an important trait required
for a robotic system to achieve autonomy. The task of object recognition involves
understanding and labeling the different components in a robot's environment.
This task becomes complicated for robots that operate in unstructured natural environments,
like forests or deep sea, due to noise in sensor measurements. Noisy sensor
measurements can potentially a ect a robot's perception of the world. To avoid being
misled by corrupted measurements, robots need to possess robust object recognition
capabilities that can handle noise in sensor measurements. Such robust object recognition
capabilities are valuable for processing large natural image datasets. One such case
of image datasets are the underwater imagery data gathered by marine scientists and
oceanographers; there, automatic object recognition capabilities could be invaluable.
Such a capability would enable the automatic analysis of these datasets to understand
natural phenomena, for instance to recognize different organisms of interest. Sifting
through such big datasets, which can range into millions of images, and making inferences
based on this data, is evolving into one of the biggest challenges in the field
research community. This motivates the need for automated object recognition and
image analysis tools. ☐ This dissertation focusses on object recognition techniques capable of operating
in noisy natural environments. A series underwater object recognition problems have
been solved as means to validate the developed object recognition algorithms. Each
technique was developed to complement the shortcomings of the existing tools available
to the research community. At first, eigen-value based shape descriptors were tasked
to solve a submerged subway car recognition problem. Despite being successful in
solving this problem, the eigen-value shape descriptor method cannot leverage textural
cues for object identification. This primary drawback, among other shortcomings, lead
to the development of a multi-layered object recognition architecture. This multilayered
architecture was tested on an scallop enumeration problem. 60-70% of scallop
instances were successfully identified. To improve the machine learning classifier of
this multi-layered framework, and also to minimize false positives, a multi-view object
classification approach is proposed. This multi-view approach combines histogram-based
global cues from a series of images of a target, captured from different heights,
to construct a machine learning classifier. This multi-view method was successful in
classifying all specimens in the available dataset. In addition to the developed object
recognition methods, a low cost ROV, named CoopROV, was designed for underwater
data collection to support research experiments.