Modified AlexNet for Dense Crowd Counting

Abstract

This paper presents a modified AlexNet to estimate the number of people in still images. The sizes of people’s heads in images differ greatly due to some factors such as perspective effect and image resolution. Feature maps with different receptive fields which are adaptive to different sizes of objects are used in this crowd counting structure. We utilize parts of AlexNet to extract features, thus getting well-trained parameters to initialize the Convolutional Neural Network. Then feature maps in different convolutional layers are merged for further feature extraction. Since the feature maps are with different sizes of receptive fields, the network is more adaptive to diversity in people’s head-size. Experiments conducted on Shanghaitech dataset demonstrate the effectiveness of the proposed method.