Improved pedestrian detection using motion segmentation and silhouette orientation

Pedestrians are the most interesting as well as vulnerable entity from both safety and security perspective in the field of video surveillance. In this article, we present a framework to detect pedestrians across a stationary camera view. Our propositions thrust upon developing a motion segmentation module and a feature extraction module for human localization. In the first stage, a background subtraction method is proposed to collect the initial set of moving objects in the processed frame. A shape descriptor is then presented to encode the pattern of human body in terms of silhouette orientation histogram. Moreover, the principle of Golden ratio is employed to formulate a part-based detector to alleviate the problem with occlusion. Both the above modules are first validated separately, and then as a unified unit, using various statistical measures. The proposed background subtraction module is simulated on twenty video clips taken from three benchmark datasets. The efficacy of our shape descriptor is validated on various image-windows taken from three publicly available datasets. The unified framework including both the modules are tested on two standard surveillance datasets. All the experimental results are uploaded at: https://sites.google.com/site/PedestriansInMotion.