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.
