Feature-based image patch classification for moving shadow detection

  • Mosin Russell

Western Sydney University thesis: Doctoral thesis

Abstract

Moving object detection is a first step towards many computer vision applications, such as human interaction and tracking, video surveillance, and traffic monitoring systems. Accurate estimation of the target object's size and shape is often required before higher-level tasks (e.g., object tracking or recog nition) can be performed. However, these properties can be derived only when the foreground object is detected precisely. Background subtraction is a common technique to extract foreground objects from image sequences. The purpose of background subtraction is to detect changes in pixel values within a given frame. The main problem with background subtraction and other related object detection techniques is that cast shadows tend to be misclassified as either parts of the foreground objects (if objects and their cast shadows are bonded together) or independent foreground objects (if objects and shadows are separated). The reason for this phenomenon is the presence of similar characteristics between the target object and its cast shadow, i.e., shadows have similar motion, attitude, and intensity changes as the moving objects that cast them. Detecting shadows of moving objects is challenging because of problem atic situations related to shadows, for example, chromatic shadows, shadow color blending, foreground-background camouflage, nontextured surfaces and dark surfaces. Various methods for shadow detection have been proposed in the liter ature to address these problems. Many of these methods use general-purpose image feature descriptors to detect shadows. These feature descriptors may be effective in distinguishing shadow points from the foreground object in a specific problematic situation; however, such methods often fail to distinguish shadow points from the foreground object in other situations. In addition, many of these moving shadow detection methods require prior knowledge of the scene condi tions and/or impose strong assumptions, which make them excessively restrictive in practice. The aim of this research is to develop an efficient method capable of addressing possible environmental problems associated with shadow detection while simultaneously improving the overall accuracy and detection stability. In this research study, possible problematic situations for dynamic shad ows are addressed and discussed in detail. On the basis of the analysis, a ro bust method, including change detection and shadow detection, is proposed to address these environmental problems. A new set of two local feature descrip tors, namely, binary patterns of local color constancy (BPLCC) and light-based gradient orientation (LGO), is introduced to address the identified problematic situations by incorporating intensity, color, texture, and gradient information. The feature vectors are concatenated in a column-by-column manner to con struct one dictionary for the objects and another dictionary for the shadows. A new sparse representation framework is then applied to find the nearest neighbor of the test image segment by computing a weighted linear combination of the reference dictionary. Image segment classification is then performed based on the similarity between the test image and the sparse representations of the two classes. The performance of the proposed framework on common shadow detec tion datasets is evaluated, and the method shows improved performance com pared with state-of-the-art methods in terms of the shadow detection rate, dis crimination rate, accuracy, and stability. By achieving these significant improve ments, the proposed method demonstrates its ability to handle various problems associated with image processing and accomplishes the aim of this thesis.
Date of Award2019
Original languageEnglish

Keywords

  • image processing
  • digital techniques
  • machine learning
  • shades and shadows
  • computer vision

Cite this

'