The detection of moving objects plays an important role in many image processing and computer vision applications

such as object recognition and tracking

as well as traffic surveillance. However

moving cast shadows are usually misclassified as moving objects because the moving shadows not only have the same movement as the moving objects but also have different intensity values with the background regions. It is well known that chromaticity and texture are the two most commonly used features in moving shadow detection. In this paper

by combining these two features reasonably

we propose a novel approach based on chromaticity and texture invariance to remove the shadow regions from the moving detection results. In our method

the moving shadows are detected not only at the pixel level

but also at the region level. When a background pixel is covered by shadow

its intensity value decreases

but the chromaticity remains the same. Therefore

we first use the chromaticity invariance at the pixel level to obtain candidate shadow regions. Chromaticity is usually used in HSV color space

which can provide a natural separation between chromaticity and luminosity. In our method

an approach applying chromaticity directly in the RGB color space is proposed

which allows for an algorithm withfewer parameters. In order to ensure that all the real shadows are included in the initial candidate shadow regions

a corresponding threshold is set conservatively in this step

which will lead to parts of the moving objects wrongly detected as shadow regions as well. To remove them from the candidate shadow regions

we first segment the candidate shadow regions into several sub-regions according to the color information. Since the moving objects and their shadows usually have different intensity values

this strategy can effectively separate the shadows from the moving objects. Then we introduce a quantitative index inspired by the local binary pattern (LBP) to measure the texture invariance of each sub-region. As regions under shadow tend to remain their textural characteristics

the real shadow regions can be finally detected in the region level. We choose five videos from a public dataset

which is most commonly used in moving shadow detection

to verify the effectiveness of the proposed method. The selected videos contain different scenes

indoor and outdoor

and the size of the shadow changes in a wide range. First

experiments to confirm the assumptions in this paper are conducted. Then

we compare our results with some classical and popular approaches via several objective evaluation criteria. The methods used for the comparison are chromaticity method

geometry method

physical method

small-region texture method

and large-region texture method. Experimental results demonstrate that our method can obtain good detection results and outperform other methods in terms of subjective visual perception and objective evaluation criteria. There are three main contributions of this paper. First

we present an approach

which directly applies the chromaticity invariance in RGB color space

which makes the number of parameters less than traditional methods using the HSV color space. Second

a quantitative index inspired by the LBP is designed to measure the texture similarity between the detected candidate regions and the corresponding background regions. Compared with conventional methods that use histogram-matching approaches

the proposed method is more accurate. Third

we provide a good detection example by combining the pixel level-based method and region level-based method. In this method

color information and texture information are both fully exploited. Experimental results demonstrate that our method can successfully detect the moving shadows in various categories of scenes.