A hybrid approach to crowd density estimation using statistical leaning and texture classification

Abstract

Crowd density estimation is a hot topic in computer vision community. Established algorithms for crowd density estimation mainly focus on moving crowds, employing background modeling to obtain crowd blobs. However, people’s motion is not obvious in most occasions such as the waiting hall in the airport or the lobby in the railway station. Moreover, conventional algorithms for crowd density estimation cannot yield desirable results for all levels of crowding due to occlusion and clutter. We propose a hybrid method to address the aforementioned problems. First, statistical learning is introduced for background subtraction, which comprises a training phase and a test phase. The crowd images are grided into small blocks which denote foreground or background. Then HOG features are extracted and are fed into a binary SVM for each block. Hence, crowd blobs can be obtained by the classification results of the trained classifier. Second, the crowd images are treated as texture images. Therefore, the estimation problem can be formulated as texture classification. The density level can be derived according to the classification results. We validate the proposed algorithm on some real scenarios where the crowd motion is not so obvious. Experimental results demonstrate that our approach can obtain the foreground crowd blobs accurately and work well for different levels of crowding.