Crowd Counting in a Highly Congested Scene using Deep Augmentation Based Convolutional Network
9 Pages Posted: 24 May 2019
Date Written: March 15, 2019
In this Paper, we proposed a unique two-level data augmentation-based approach for crowd counting in a still image. Most of the existing researchers have used deep and shallow two different types of networks for high and low-level feature extraction. We used a single framework for high and low-level feature extraction. In our approach, we first convert our training samples to polar coordinates and use the magnitude of the resultant image and the skeleton of the image to feed directly to the deep convolutional neural network with the original training samples. Most of the crowd datasets have very less amount of training samples and deep learning-based approaches require large amounts of training data. This approach of two-level augmentation helps our deep network to learn high-level features directly from original samples as well as the low-level features from augmented samples. This feature extraction is then followed by a fully connected regress network for local patch count. We used two public datasets to demonstrate our proposed approach, UCF, and Shanghaitech datasets. The experiments show that our deep augmentation-based approach outperforms most of the state of the art methods.
Keywords: Deep Augmentation, Convolutional Neural Network, Crowd Counting
Suggested Citation: Suggested Citation