Real-time extraction of maximally stable extremal regions on SoC for visual surveillance

Abstract

Numerous techniques and algorithms have been developed for object detection, tracking and recognition. Most of thepublished techniques and algorithms are primarily software-based. Yet, few attempts have been made to implement some of thealgorithms in hardware, and those attempts didn't yield the optimal results in terms of accuracy, power and memory requirements. Furthermore, most of thepublished hardware techniques are not suitable for real-time applications. In this thesis, a novel low power real-time architecture is developed to efficiently implement the Maximally Stable Extremal Regions (MSER) algorithm. The proposed architecture memory requirement is 18% less than that of the state-of-the-art hardware implementation of the MSER algorithms published in the literature. Also, the architecture is very suitable for ASIC (Application Specific Integrated Circuit) due to the low memory requirement and the use of state forward sequential state machine controller. In addition to the architecture, a novel algorithm for parallel Maximally Stable Extremal Regions (MSERs) detection is presented. The classical MSER algorithm makes use of the intensity image and the inverted intensity image to detect dark and bright MSERs, respectively, in two separate runs. The proposed algorithm uses a novel labeling technique to provide equivalent detection results, to those of the classical algorithm, in a single run based on only either one of the intensity image versions. As a result, the new parallel MSER detector algorithm requires reduced memory accesses by nearly half and achieves nearly 50% savings on the overall MSER computations. Consequently, the proposed parallel MSER algorithm has a twofold faster execution time, doubling the FPS (frames/second) rate. The results of the intensive testing using MATLABproved that the proposed parallel MSER detector is more accurate than MATLAB's built-in implementation of the classical MSER detector. The enhanced performance of the proposed algorithm through parallelism will enable more MSER based applications. We further extended the MSER algorithm detection capability to depth information by introducing the XMSER detector that shows a robust detection performance for both single and multiple static and dynamic object detection and tracking. Finally, as illustrative potential applications for the MSER detector, two novel biomedical imaging applications are introduced, namely, the MERSCoV (Middle East Respiratory Syndrome Coronavirus) Bio-surveillance system and the MRI (Magnetic Resonance Imaging) Browser for cancer detection.

Date of Award	2014
Original language	American English
Supervisor	Hani Saleh (Supervisor)