Crowd Counting via Weighted VLAD on Dense Attribute Feature Maps

摘要

Crowd counting is an important task in computer vision, which has manyapplications in video surveillance. Although the regression-based framework hasachieved great improvements for crowd counting, how to improve thediscriminative power of image representation is still an open problem.Conventional holistic features used in crowd counting often fail to capturesemantic attributes and spatial cues of the image. In this paper, we proposeintegrating semantic information into learning locality-aware feature sets foraccurate crowd counting. First, with the help of convolutional neural network(CNN), the original pixel space is mapped onto a dense attribute feature map,where each dimension of the pixel-wise feature indicates the probabilisticstrength of a certain semantic class. Then, locality-aware features (LAF) builton the idea of spatial pyramids on neighboring patches are proposed to exploremore spatial context and local information. Finally, the traditional VLADencoding method is extended to a more generalized form in which diversecoefficient weights are taken into consideration. Experimental results validatethe effectiveness of our presented method.