Shadow removal utilizing multiplicative fusion of texture and colour features for surveillance image

Abstract

Automated surveillance systems often identify shadows as parts of a moving object which jeopardized subsequent image processing tasks such as object identification and tracking. In this thesis, an improved shadow elimination method for an indoor surveillance system is presented. This developed method is a fusion of several image processing methods. Firstly, the image is segmented using the Statistical Region Merging algorithm to obtain the segmented potential shadow regions. Next, multiple shadow identification features which include Normalized Cross-Correlation, Local Color Constancy and Hue-Saturation-Value shadow cues are applied on the images to generate feature maps. These feature maps are used for identifying and removing cast shadows according to the segmented regions. The video dataset used is the Autonomous Agents for On-Scene Networked Incident Management which covers both indoor and outdoor video scenes. The benchmarking result indicates that the developed method is on-par with several normally used shadow detection methods. The developed method yields a mean score of 85.17% for the video sequence in which the strongest shadow is present and a mean score of 89.93% for the video having the most complex textured background. This research contributes to the development and improvement of a functioning shadow eliminator method that is able to cope with image noise and various illumination changes.