PeopleNet: A Novel People Counting Framework for Head-Mounted Moving Camera Videos

Abstract

Traditional crowd counting (optical flow or feature matching) techniques have been upgraded to deep learning (DL) models due to their lack of automatic feature extraction and low-precision outcomes. Most of these models were tested on surveillance scene crowd datasets captured by stationary shooting equipment. It is very challenging to perform people counting from the videos shot with a head-mounted moving camera; this is mainly due to mixing the temporal information of the moving crowd with the induced camera motion. This study proposed a transfer learning-based PeopleNet model to tackle this significant problem. For this, we have made some significant changes to the standard VGG16 model, by disabling top convolutional blocks and replacing its standard fully connected layers with some new fully connected and dense layers. The strong transfer learning capability of the VGG16 network yields in-depth insights of the PeopleNet into the good quality of density maps resulting in highly accurate crowd estimation. The performance of the proposed model has been tested over a self-generated image database prepared from moving camera video clips, as there is no public and benchmark dataset for this work. The proposed framework has given promising results on various crowd categories such as dense, sparse, average, etc. To ensure versatility, we have done self and cross-evaluation on various crowd counting models and datasets, which proves the importance of the PeopleNet model in adverse defense of society.