People Counting in Extremely Dense Crowd

Abstract

This thesis presents a novel method for crowd counting in static images under varying crowd density conditions. Existing methods for crowd counting suffer from poor performances due to camera limitations and crowd conditions: clutter, few pixels per person, severe occlusion, camera perspective, and varying image resolutions to name a few. The proposed method seeks to overcome these challenges by adopting a patch-based approach for local attention. The input image is first fed to a Patch Classifier Network that determines the optimal patch-size for the subsequent processing. The image is then divided into non-overlapping contiguous patches. Each patch is fed to a Count Regressor Network to estimate patch count. Finally, individual patch counts are summed up to obtain the final total count. The key inspiration behind the optimal patch-size based counting is to mimic a human annotator (counter) paying close attention to a local region in the image and scanning the whole image sequentially while zooming in and out based on the local crowd density. The proposed architecture produces state of the art results on publicly available crowd datasets: UCF_CC_50, AHU-CROWD, and ShanghaiTech Part A and B.