Communications in Information and Systems
Volume 17 (2017)
Counting congested crowds under wild conditions with a multi-task Inception network
Pages: 1 – 24
Counting of congested crowds has been widely applied in surveillance event detection, public safety control, and traffic monitoring. Early studies mainly focus on designing hand-crafted features. However, counting performance based on hand-crafted features maybe easily influenced by issues such as partial occlusion, scale variation, and illumination change. Convolutional neural network (CNN) has shown great success in visual crowd counting. In this work, a multi-task Inception network is proposed for crowd counting in congested crowds. Crowd count is predicted through summing up a density map estimated by the proposed network. Three Inception blocks are employed to automatically extract multi-scale features from different patches cropped from the crowd image. The network can jointly estimate the density map, crowd density level, and background / foreground separation. Counting performance obtained through multi-task learning is superior to that obtained through only estimating density map. Contrastive evaluations based on three benchmarking datasets are implemented with several state-of-the-art CNN-based crowd counting approaches. Results indicate the accuracy and robustness of our network in counting congested crowds. The multi-task Inception network almost outperforms the state-of-the-art counting approaches in terms of mean absolute error and mean squared error.
This work has been supported by the National Natural Science Foundation of China under Grant No. 61501060 and No. 61703381, the Natural Science Foundation of Jiangsu Province under Grant No. BK20150271, Key Laboratory for New Technology Application of Road Conveyance of Jiangsu Province under Grant BM20082061708.
Published 30 January 2018