Counting Vehicles in Static Images

This work addresses the problem of counting vehicles in static images with no geometric information of the scene. Five convolutional neural network architectures were studied, implemented and trained as the main part of this work. Also, a dataset that consists of 19 310 images from 12 views that captures 7 different scenes were taken as part of this work. The trained networks map the appearance of the input sample to its corresponding vehicles density map, which can be easily translated to the vehicle count with keeping the localization of the vehicles in the input image. The main contribution of this work is in an application and a comparison of the state-of-the-art solutions to the problem of object counting. Most of them were mainly designed to count pedestrians in crowded scenes or for medicine images, so the major goal was to adapt these solutions for vehicle counting task. The implemented models were evaluated on TRANCOS dataset and large custom dataset with the GAME metric. Their performance is compared and the results are discussed.

objects counting by estimating a density map, custom carpark dataset, convolutional network, deep learning, landmarks localization, GAME metric