Crowd sensing using deep learning

Abstract

Crowd density has risen across the worlds as the human population increases across the past decades. This get worse as people tends to gather into cities and may cause a strain on certain facilities when they are being overused or cause accidents due to overcrowding. Many newer model for crowd counting task employ the density map approach. Density map approach is good for estimating people count in a very dense image but lacks interpretability in the model’s prediction and accuracy. These models were also not built with the consideration of using it for crowd surveillance tasks but more so focus on getting the best accuracy on the images. This report will go through the different deep learning models used for crowd counting and implement and train a model that deem the most fit for common crowdsurveillance purposes. It will still be evaluated against the different models and a website will be built around the model to show the use of such model in practical situations.