People Counting from Moving Camera Videos through PeopleNet Framework

Counting the people in a moving camera video or picture is difficult since the positions of the people and the camera overlap in the frame. Traditional object recognition techniques use feature matching or optical flow mechanisms, but unable to perform people estimation due Limited Viewpoints, Scale Variations, Object Appearance Variability, and Dynamic Environments issues. This research aims to enhance object counting research in a manner that differs from traditional techniques. With the aid of density map generation, this study proposed a transfer learning-based PeopleNet model that counts the correct number of people from moving camera videos. The proposed model has been developed by fine-tuning two fully connected (FC) layers at the end of the standard pre-trained VGG16 model. In this model, the top layers are frozen. There are no public or benchmark datasets available to evaluate the model’s results. Therefore, we introduced a new dataset captured by a moving camera to counter such a scarcity of resources. The PeopleNet has achieved impressive results in dense, average, and sparse crowd scenarios and has established a new standard for future people counting studies. The proposed model for counting people from moving camera videos has exhibited behavioral and favorable results. It could be useful in real-time applications.