Fake Shadow Detection Using Local Histogram of Oriented Gradients (HOG) Features

Shadows cast by the moving objects may lead to several errors in the process of moving object detection and tracking. Since the shadows are connected to the object and move along with it, false object detection may occur in addition to the problem of false connectivity and loss of background texture. Hence shadow detection is an important preprocessing step for a robust visual surveillance system. However, the conventional methods which usually use a static threshold over the color and/or intensity channels for shadow detection may fail when some object regions have properties identical to the cast shadow (fake shadow). In this project, we present a novel shadow detection and removal scheme which can effectively deal with the problem of fake shadows using the HOG (Histogram-of-Oriented Gradients) features. In the initial stage of moving object detection we make use of GMM (Gaussian Mixture Model) to properly segment the foreground regions. Since HSV color space gives a better separation of chromaticity and intensity, it helps to detect the shadows in the segmented foreground but at the same time may also misclassify some object regions as shadow regions. We exploit the fact that these object regions (misclassified regions/fake shadows) change the entire background information as compared to the cast shadows which mostly cause intensity variations over the background and perform a local feature matching process to properly identify the real shadow and fake shadow regions. Once the regions are identified it becomes easy to remove the shadows without any loss of information in the object regions. Experimental results indicate that the proposed method achieves good results in outdoor scenarios.