Real-time recognition and tracking of texture-less objects

Recognition of texture-less objects is one of the most challenging problems in the eld of object detection. This stems from the scarcity of resonant information, which impedes the formation of robust features for its description. Furthermore, as recent demands in mobile technologies have limited com...

Full description

Saved in:
Bibliographic Details
Main Author: Chan, Jacob Huat Kwang
Other Authors: Jimmy Addison Lee
Format: Theses and Dissertations
Language:English
Published: 2017
Subjects:
Online Access:http://hdl.handle.net/10356/73038
Tags: Add Tag
No Tags, Be the first to tag this record!
Institution: Nanyang Technological University
Language: English
Description
Summary:Recognition of texture-less objects is one of the most challenging problems in the eld of object detection. This stems from the scarcity of resonant information, which impedes the formation of robust features for its description. Furthermore, as recent demands in mobile technologies have limited computational resources, it is imperative to design modern detectors that are fast, compact and robust to stay relevant. Therefore, this thesis proposes a framework to tackle these issues. To establish a strong foundation for texture-less recognition, we rst introduce BORDER (Bounding Oriented-Rectangle Descriptors for Enclosed Regions). Essentially a keypoint-descriptor based algorithm, it features a highly repeatable occlusion resistant interest-point detector termed Linelets, coupled with a regional object encompassing oriented rectangle rotation scheme that is capable of attaining state-of-the-art detection rates in texture-less object recognition. Next, we take a step closer towards robust real-time texture-less detection with the introduction of BIND (Binary Integrated Net Descriptor), where lightweight binary nets are stacked in layers onto detected linelets to encourage discriminative object description through its high-spatial resolution, and unique binary matching scheme. This enables highly precise encoding of the object's edges and internal homogeneous information to attain even higher detection rates at a fraction of the cost in speed and memory space when compared to BORDER. The last hurdle towards real-time detection involves geometric veri cation, whereby keypoint matches are ltered to obtain their inliers. As widely used algorithms are highly iterative (i.e. RANSAC, Hough Transform), they are mostly too expensive for real-time applications. Therefore, we propose an approach coined F-SORT (Fast Sequence Order Re-sorting Technique), where matches are sorted into local sequence groups for geometric validation along di erent orientations to enable a signi cant reduction in computational cost, while inciting robust outlier rejection. Finally, we fuse the entire framework in a mobile application for detecting and tracking texture-less objects in real-time.