Characterizing Scattered Occlusions for Effective Dense-Mode Crowd Counting

Abstract

We propose a novel deep learning approach for effective dense crowd counting by characterizing scattered occlusions, named CSONet. CSONet recognizes the implications of event-induced, scene-embedded, and multitudinous obstacles such as umbrellas and picket signs to achieve an accurate crowd analysis result. CSONet is the first deep learning model for characterizing scattered occlusions of effective dense-mode crowd counting to the best of our knowledge. We have collected and annotated two new scattered occlusion object datasets, which contain crowd images occluded with umbrellas (csoumbrellas dataset) and picket signs (cso-pickets dataset). We have designed and implemented a new crowd overfit reduction network by adding both spatial pyramid pooling and dilated convolution layers over modified VGG16 for capturing high-level features of extended receptive fields. CSONet was trained on the two new scattered occlusion datasets and the ShanghaiTech A and B datasets. We also have built an algorithm that merges scattered object maps and density heatmaps of visible humans to generate a more accurate crowd density heatmap output. Through extensive evaluations, we demonstrate that the accuracy of CSONet with scattered occlusion images outperforms over the state-of-art existing crowd counting approaches by 30% to 100% in both mean absolute error and mean square error.