Research On Detecting Moving Objects Using Background Subtraction And Frame Difference

There has been a significant development in the area of computer vision in the lastdecades. Detecting regions that correspond to moving objects such as people and vehicles invideo is the first basic step of almost every vision system. The accuracy of the outputs fromthis stage is crucial for the subsequent vision processing. Unfortunately fast and reliablemotion segmentation scheme that works in all kinds of view is an open and difficult problemdue to dynamic changes in natural scenes.Most approaches reported in the literature use background subtraction techniques for thedetection of moving objects in video sequences. There are many challenges in developinggood algorithm for detecting moving objects. Our main concern focuses on solving problemssuch as continuous and sudden light changes, temporal and permanent variation inbackground, shadow noise elimination et al. Therefore, a combined algorithm of backgroundsubtraction and frame difference has been proposed in this work, which uses both spatial andtemporal information for the detection of moving objects.At first, the background image is modeled using robust statistical properties of pixelsfrom successive frames during the training time, which is an important step for thebackground subtraction. The performance of frame difference method for detecting movingobjects was analyzed and it was found that the method is fast at adapting dynamicillumination change. Hence with the aim of overcoming sudden lighting change, we haveincorporated it to the proposed algorithm in order for the reference image remains adaptive tonatural scene change. The moving object is then extracted by subtracting the estimatedbackground image from the current frame. An automatically generated threshold rule isapplied to the background subtraction result in order to get potential foreground pixels. Thisgives our algorithm a powerful advantage and enables it to manage with most severeilluminations and lighting change conditions.Shadows and transient noise associated with the background subtraction result wereanalyzed and tackled. The presence of a shadow was first hypothesized based on some initialevidence that the photometric information for all pixels marked as foreground was evaluated.Then texture invariant property between pixels and corresponding background were analyzedfor all the potential shadow regions. The proposed approach is adaptive, because it allows the background model to be updated.The experimental results conducted on several indoor and outdoor sequences confirm theperformance and accuracy of the proposed algorithm in different illumination conditions.