Pedestrian group attributes detection in crowded scenes

Recently the traditional video surveillance systems of crowd scenes have been deployed in various areas of applications; health monitoring, security etc. Monitoring crowds and identifying their behaviors is one of the most interesting applications of visual surveillance as it is very difficult to assess crowds by human experts. In this paper, we present inter-group and intra-group properties of crowd scene; namely, we investigated collectiveness, stability, uniformity and conflict properties of crowds. A collective transition algorithm is used for crowd scene detection and segmentation. Based on this algorithm, a set of visual descriptors are extracted to quantify the group properties. The descriptors convey deeper scene information and can be effectively deploy in large crowd scene. Experiments on hundreds of crowd scenes videos were carried out on publicly available datasets. Quantitative evaluation shows that linear SVM display superior accuracy, precision, recall and F-measure in classifying human behaviors when compared to a k-nearest neighbor (kNN), and Decision Tree (DT) classifiers.