Robust and Efficient Corner Detection for Computer Vision

Abstract

Corner detection is a fundamental computer vision problem that has been widely studied in image retrieval, object tracking, motion estimation, visual localization and 3D reconstruction. The accuracy and repeatability of corner detection are important for image matching and retrieval, while the detection accuracy and localisation ability are vital for visual localisation, motion estimation and 3D reconstruction. For real-time computer vision tasks, it is important to achieve fast corner detection with high detection accuracy and repeatability.

This thesis addresses the problem of corner detection and localisation with high accuracy and repeatability. In order to improve the detection performance of the current state-of-the-art in corner detection, an improved shearlet transform and a novel complex shearlet transform are proposed to overcome the problems of traditional shearlets and achieve better localisation of distributed discontinuities, particularly with the ability to extract phase information from geometrical features. Moreover, a multi-directional structure tensor and a multi-scale corner measurement function are proposed to make full use of the structural information from the improved shearlets for detection, and a new rotary phase congruence tensor is proposed to utilize all amplitude and phase information of the complex shearlets for detection. As a result, two new corner detectors are proposed. Experimental results demonstrate that their localisation ability and detection accuracy are superior to current detectors, and their repeatability is generally higher than current corner detectors as well as recent deep learning based interest point detectors. Therefore, they have great potential for applications in computer vision tasks.

To meet the needs of real-time computer vision tasks, especially real-time portable tasks, a new type of filter that can enhance corners and suppress edges as well as noise simultaneously is proposed to simplify the detection architecture and improve its parallel computing performance, and a novel corner detector with high computational efficiency is proposed. The corresponding field programmable gate array (FPGA) design is provided as well. Experimental results show that, with very low computational cost and simple architecture, the proposed detector can achieve similar detection accuracy and repeatability of current corner detectors while being potentially useful for real-time computer vision applications.