Some shadow detection and removal techniques have been proposed in recent years. Zhang et al. [1] classify these techniques into four categories: color model, statistical model, textural model, and geometric model.
The principle of color model is that by observing or finding the color change between the shaded and non-shaded pixel. Cucchiara et al. [2] used HSV color space to remove moving shadow. The concept is that the hue component in shaded pixel would remain roughly the same comparing to the pixel is non-shaded. And the saturation component would decrease. Some researchers proposed shadow detection methods based on RGB color space and normalized-RGB color space. Yang et al. [3]
described the ratio between a pixel in shaded region and its neighboring shadow pixel in current image would close to those in the background image. For example, in current image, a pixel at (x, y) is a shaded pixel, and it neighboring pixel, the pixel at (x+1, y), is also a shaded pixel. The intensity ratio of these two pixels will be equal to the intensity ratio of two pixels at same coordinate in background image. Besides, another feature they have used is that the change of normalized r and g channel between current and background image would change slightly. Cavallaro et al. [4]
found that the color components do not change their order and photometric invariant features do not change their value a lot when a shadow occurs. They firstly selected some candidates of shadow region. By spatial and temporal verifications, they could eliminate some shadow candidates that were detected by mistakes. However, as mentioned before, the format of video sequence may not be colorized or the computation loading usually increase.
Besides color model, some authors also utilized statistical model in their
a pixel belongs to shadow or not. Zhang et al. [1] led in an illumination invariance feature and then analyzed and modeled shadow as a Chi-square distribution. They classified each moving pixel into shadow or foreground object by performing a significance test. Song et al. [5] exploited Gaussian model to represent the constant RGB-color ratios, and by setting ±1.5 standard deviation as a threshold to discriminate a moving pixel which belongs to shadow or foreground object. Nicolas et al. [6] proposed GMSM (Gaussian Mixture Shadow Model) for shadow detection.
The GMSM was integrated into a background detection algorithm based on GMM.
They test if the mean of a distribution could describe a shaded region, and they will select this distribution to update corresponding Gaussian mixture shadow model. But their method should require a lot of memory, and the computation loading is also a little heavy.
The idea behind the texture model is that the texture of the foreground object would totally differ from the texture of background at the same position, but the texture would be the same inside the shaded region. Joshi et al. [7][8] proposed an algorithm that can learn and detect shadow by using support vector machine. They defined four image features, including intensity ratio, color distortion, edge magnitude distortion and edge gradient distortion. By using two SVM classifiers, they led in co-training architecture and make these two classifiers can help each other in training process. A small set of shadow labeled samples need to be inputted before training SVM classifiers. Although this method just requires small set of shadow labeled samples, it is still inconvenient to provide such shadow labeled samples for different video sequences. Leone et al. [9] presented a shadow detection method by using Gabor features. But it is a little computationally inefficient. Mohammed et al. [10]
proposed their method by using division image analysis and projection histogram
frame, and it can highlight homogeneity property of shadows. After taking an adaptive threshold, they used both column and row projection histogram analyses to eliminate the left pixels which locate at boundary of shadow.
Benedek et al. [11] proposed a method that uses LUV color model. They used
“darkening factor”, distortion of U and V channels and microstructural response as determinative features. Microstructural response represents a local texture feature.
The authors modeled these features by Gaussian model. By calculating the probabilities of background, shadow, foreground and taking threshold, their proposed algorithm could tell foreground objects, background and shadow apart. Xiao et al. [12]
proposed a shadow removal method which based on edge information for traffic scenes. They applied an edge extraction technique and then used morphological operations to remove the boundary of shadow. Then, in order to cope with car occlusion problem which arose from shadow, the authors exploited the spatial property to separate occluded cars. Finally, they reconstructed the size of each object and obtained real shadow regions. However, due to the property of this method, if the region inside the shadow has texture, for example, including lane marking etc., then this method could be failed. Besides, considering the occlusion problem which caused by shadow, if the occlusion situation is complicated, for example, the shape of occluded cars is concave, the separation method that using spatial property would also not take effect.
Geometric model attempts to use object geometry, the information that could be obtained from ground surface to eliminate shadow regions or its effect. Hsieh et al.
[13] analyzed vehicle histogram and calculated the lane center. Then, the lane dividing lines can be detected. They developed a horizontal and vertical line-based method that could eliminate shadow according to these lane dividing lines. However,
Utilizing color information for shadow removal may have good result, if develop methodology properly. But, unfortunately, not all application systems suit to use color camera. Besides, the efficiency of this kind of methods is usually not satisfied.
Statistical method is easy to lead in developing shadow removal method. However, it usually requires manually shadow labeled samples for training. Texture model can have better result when illumination of scene is not stable. In addition, this kind of method doesn’t require color information. But, if the object is textureless, texture model may not have good performance. Geometric model usually fits specific scene due to it depends on geometric relations of object and scene. By considering different characteristics of these methods, we decide to utilize texture and statistical model to achieve moving shadow removal. We hope our proposed method is stable and without using color information by applying texture model. And, we utilize statistical method to enhance performance and deal with the textureless problem.