Adaptive Background Modeling and Foreground Detection in Video Sequence Using Artificial Neural Network
N. Satish Kumar · 2012
This paper presents an Artificial Neural Network (ANN) approach of using the texture properties to model the background and detect the moving objects in video sequence. Extraction of foreground object is a prime step of information extraction in many computer vision applications. The proposed approach is adaptive to the dynamic background. Besides the usefulness of being able to segment moving and non moving objects in video sequence, detecting moving objects provides provision for recognition, classification and activity analysis. Each pixel is modeled using Local Binary Pattern (LBP) calculated along a circular region of r radius over a pixel. Then Background model is build using Self Organizing Map ANN and the next incoming frames are compared against the background model using Mahalanobis distance to decide whether it belongs to background or foreground. The proposed method can handle scenes containing moving background, gradual illumination variations. The algorithm proposed has been tested on some videos taken by stationary camera and compared with the existing algorithms in terms of detection accuracy and processing speed for color video sequence which is critical for color video surveillance system. It was found that our approach resulted in 93% of Detection Accuracy and on average 1.5 times faster than existing algorithms. technique used. Background modeling technique should be capable of dealing with movement through cluttered areas, objects overlapping in the visual field, shadows, lighting changes etc (1). In this paper we explore the traditional approaches based on background modeling methods which typically fail in general situations. We also propose a robust, machine learning system that is flexible enough to handle variations, in lighting, moving scene clutter, multiple moving objects and other arbitrary changes to the observed scene. The detection of moving objects in video streams is the first relevant steps of information extraction, and can be used for recognition, classification, and activity analysis more efficiently, since only moving pixels need to be considered.