Summary: Clutter and occlusions hinder conventional crowd density estimation algorithms from yielding accurate results for different crowd density levels. A hybrid approach is presented here by combining local and global features. First, preprocessing of the input images is used to reduce the background noise. Then, the ratio of foreground blobs to the whole image is calculated with a mechanism for threshold segmentation. Finally, a regression algorithm based on the local features is used to analyze images below the threshold with a classification algorithm based on global features images used to analyze images above the threshold. A texture descriptor combining wavelet transforms and a gray level co-occurrence matrix is then used to improve the classification accuracy for the classification algorithm based on the global features. Tests demonstrate that this method is both accurate and robust for different crowd density levels.