Output details

Image Processing: Object Segmentation Using Full-Spectrum Matching of Albedo Derived from Colour Images

<23> Significance:

This patent describes a ground-breaking object detection algorithm (‘Spectral 360’®) which is resilient to illumination-induced colour change in videos, and to other sources of error. It meets the needs of video analysis for challenging imaging settings, such as visual surveillance, which needs reliable and rapid processing. For example, Spectral 360® was used by police investigators, enabling them to remarkably shorten investigations and significantly reduce investigation cost. It led to the creation of AVA Technologies Ltd. It is a significant underpinning for the ‘Adaptive Video Analytics Software’ impact case study, in this submission.

Research process / content:

Spectral 360® was developed from rigorous doctoral research, including a comprehensive literature review, algorithm development, and a rigorous methodology revolving around a systematic performance assessment, anchored on experimental designs with qualitative and quantitative performance measures, including tests of statistical significance. It is underpinned by a mathematical formulation of physics models for the relevant imaging conditions.

Moreover, Spectral 360® has been tested rigorously against 22 other algorithms from all over the world, on a comprehensive benchmarking set of challenging videos, using seven performance measures (Precision; F-Measure; Percentage of Wrong Classifications; False Negative Rate; False Positive Rate; Specificity; Recall). The test dataset consists of 31 real-world videos (over 80,000 frames); in 6 categories of challenges (simple; dynamic background; heavy camera jitter; objects moving intermittently; hard and soft shadows, intermittent shades; thermal camera videos). Spectral 360® emerged as the clear winner against algorithms from Technische Universität Berlin, Technische Universität München, Queen Mary University of London, Fraunhofer IOSB, intuVision Inc., National Research Center of Italy, University of Amsterdam, ICAR-CNR, Kyushu University, Brunel University, State University of New Jersey, Mitsubishi Electric Research Laboratories, Massachusetts Institute of Technology, École Polytechnique de Montréal, University of Washington, Université de Sherbrooke. (see http://www.changedetection.net/ [Accessed on 11/11/2013])