Abstract: Crowd density estimation plays a central role in intelligent monitoring.Deep neural network usually outperforms conventional approaches based on manual features owing to its data-driven superiority.However,deep neural networks are still far from optimal solution because of the scarceness of large-scale datasets.To address this problem,this paper investigated the feasibility of several solutions which are training shallow neural network from scratch,using fullyconnected layer features of pretrained deep neural network and aggregating convolutional features by way of fisher vector(FV).Aiming at the problem of unbalanced distribution,this paper further proposed several classification evaluation criteria.Comprehensive experiments were carried out on benchmark PETs2009 dataset.Results show that convolutional features outperform existing hand-crafted ones.Moreover,utilizing deep convolutional features based on transfer learning usually leads to better performance than the models trained from scratch.Finally,simpler pretrained models such as AlexNet can generalize better mobility of the lower layer features than more complicated ones such as VGGNet.

Key words: Crowd density estimation, Deep convolutional neural network, SVM, Texture feature, Transfer learning