Crowd Counting with Dilated Inception Convolution

Abstract

Convolutional neural network (CNN) has been successfully applied to image-based crowd density estimation. However, large computational resources are required in previous CNN-based methods. Therefore, to overcome these drawbacks, this paper proposes a lightweight crowd density map estimation architecture with Dilated Inception Convolution Neural Network (DICNN). The proposed method not only extracts scale-aware informative features, but also effectively reduces the number of parameters of the CNN architecture. In addition, the proposed method is trained along with DICNN in an end-to-end fashion via both pixel-wise Euclidean distance and density-level relevant (DLR) loss for global optimization. Extensive experiments on several publicly available datasets have shown that the proposed method outperforms state-of-the-art approaches in almost all datasets with far fewer parameters.