Boosting deep attribute learning via support vector regression for fast moving crowd counting

Crowd counting has recently attracted extensive attention in research. However, the existing research mainly focuses on investigating crowd counting of static or slow moving crowd estimating, while fast moving crowd counting is left unexplored. The fast moving crowd counting is indeed extremely important for urban public safety management. In this paper, we propose a novel more effective fast moving crowd counting algorithm. The proposed approach utilizes support vector regression and spatial-temporal multifeatures to boost deep cumulative attribute learning. To this end, first a novel spatial-temporal multifeature is proposed by joining super-pixel based multi-appearance features and multi-motion features to solve fast moving crowd counting. Second, a novel deep accumulated attributes learning architecture is proposed based on very deep learning architecture VGG16. Third, a novel boosting deep attribute Learning algorithm is proposed based on late fusion of proposed deep cumulative attribute learning and proposed spatial-temporal multi-features based support vector regression for improving predication performance of deep learning. We perform corresponding experiments on three public datasets including UCSD dataset, PEST2009 dataset and Mall dataset. The experimental results demonstrate that proposed Boosting DAL-SVR method is effective to cover the shortage of deep learning in solving regression problems. Meanwhile it demonstrates that proposed Boosting DAL-SVR is more effective and robust rather than other state-of-the art methods for fast moving crowd counting problem. (C) 2017 Elsevier B.V. All rights reserved.