Human shadow detection for real-time applications

This paper presents a new and effective method for detecting cast shadows of humans for real-time applications. The method proposes a new model of pixel-gradient magnitude and direction to detect human shadows in most possible scenarios occurring in videos. First, the candidate shadow regions are produced using intensity measurements of the sub-divided image blocks. A new model of pixel-gradient magnitude is then used to find the best matched weight between each block in the given frame and the corresponding background. Finally, the direction of intensity ratio between given frame and background frame is computed and compared with the best-matched block to find those having similar gradient directions. Using this technique improves the classification result in many environmental problems including, camouflages, non-textured surfaces, and chromatic shadows. The quantitative results on three challenging datasets show a significant performance improvement of the proposed method over two well-known shadow detection methods.