People Count from Surveillance Video Using Convolution Neural Net

Abstract

People counting is used to count the quantity of people in the picture. People counting is not an easy task if it is done manually by our hand because we can lost count in the middle of doing this laborious task, especially when dealing with object that intersects with each other or dense crowd. This project automates the counting process by building a machine learning system that can convert a video into frames, then the model will output number of objects in a particular frame. We built the model using convolutional neural network (CNN) technique. The system that we built is capable of counting pedestrians in a mall. The frames/images are generated from CCTV that is placed somewhere in the mall. From those frames/images, the system will output how many pedestrians at that particular place in the mall. VGG16 is used to excerpt the topographies of the image and structural similarity index (SSIM) for measuring the similarity among the given images. Then, use the similarity measure as a loss function, named Euclidean and local pattern consistency loss. The experimental results show the predicted number of people and exact number of people in the image with 90% of accuracy using convolution neural net.