EstimatingLi, MengChen, TaoLi, ZhihuaLiu, Hezi count and density maps from crowd images has a wide range of applications such as video surveillance, traffic monitoring and pedestrian dynamics study. The state-of-the art deep learning approaches generally build a multi-column deep network architecture. Although such architectures perform well, the inference cost is neglected. In this paper, we apply the net-work compression to the convolutional neural network-basedConvolutional Neural Networks (CNN) crowd density estimationDensity estimation model, to reduce its storage and computation costs. Specifically, we rely on using l1-norm to select unimportant filters and physically prune them. Convolutional filters with small scaling factor value and its corresponding kernels in the next layer will be removed. The model can be trained to improve its regression performance and identify the unimportant filters at the same time. A state-of-the art model, the CSRNet, is tested in the ShanghaiTech dataset. Our method can reduce inference costs for the network significantly (up to approximately 78%) while regaining close to the original accuracy by retraining the networks.
An Efficient Crowd Density Estimation Algorithm Through Network Compression
Springer Proceedings Phys.
2020-11-17
9 pages
Article/Chapter (Book)
Electronic Resource
English
Density estimation , Convolutional neural networks , Network compression Physics , Soft and Granular Matter, Complex Fluids and Microfluidics , Transportation Technology and Traffic Engineering , Complex Systems , Computer Appl. in Social and Behavioral Sciences , Statistical Physics and Dynamical Systems , Physics and Astronomy
An Efficient Crowd Density Estimation Algorithm Through Network Compression
British Library Conference Proceedings | 2020
|On crowd density estimation for surveillance
IET Digital Library Archive | 2006
|Higher-order SVD analysis for crowd density estimation
British Library Online Contents | 2012
|