Partially Occluded Object Detection in Video Sequences

Abstract

Object detection algorithms play an important role in enabling machines to understand visual information from images or video sequences. Background Subtraction is one of the object detection methods that is used to separate the foreground objects from the background scene for accurate moving object detection. However, occlusion presents a challenge for Background Subtraction methods in which an object is hidden by other objects, and limited features of objects are visible, leading to inaccurate localization of the objects. The work presented in this thesis is to solve the partial occlusion problem. newline The first contribution is an adaptive background subtraction and updation method by using a two-level thresholding approach to improve pixel-wise segmentation. The background subtraction model is initialized to generate an initial threshold value, followed by updating the background subtraction model to detect subsequent changes in the frames. The performance of the proposed background subtraction method has been evaluated on two benchmarked datasets: Highway and PETS2006. Experimental results indicate that the proposed method produces a good Precision of 0.9496 for the Highway dataset and 0.9254 for the PETS2006 dataset. It was also observed that the F-Measure value is 0.9449 and 0.9247 for both datasets respectively representing correct segmentations of the pixels. There are few frames associated with shadow leading misclassification of pixels. newlineThe second contribution is to identify and eliminate shadow pixels from a segmented frame by proposing an intensity ratio-based YCbCr color space channel. The method involved computing an intensity ratio between the Y color channel and the inverted segmented mask to detect the shadow key points, followed by a color correction method to remove the shadow pixels.