Convolutional Feature Based Unbalanced Crowd Density Estimation

More and more people attend public activities such as celebration and assembly,leading to some public safety issues,which receives increasing attentions.Due to no enough aware and precaution to much high crowd density level in local regions,some stampede have happened in public activities such as the stampede happened in Shanghai bund and school named Mingtong in Kunming,Yunnan.To avoid the similar issues,a large scale public monitoring system should be established,which can recognize crowd density level automatically and give a alarm when some local regions have higher crowd density than the threshold level,showing significance for public safety.The existing crowd density estimation methods can be divided into pixel feature based methods and texture feature based methods.Pixel feature based methods are simple and effective but are inaccurate to those scenes with complex background,occlusion and high crowd density.In contrast,texture feature based methods show good performance in high crowd density scenes while perform poor in low crowd density scenes as perspective,uneven crowd distribution and complex background.A deep-learning-based method is proposed for crowd density estimation in this thesis.Deep neural network usually outperforms conventional approaches owing to its data-driven superiority and powerful representation.However,deep neural networks are still far from optimal because of the scarceness of large-scale datasets.To address this,we investigate the feasibility of several solutions: training shallow neural network from scratch,estimating density based on fully-connected features from deep neural network and aggregating convolutional features by way of Fisher Vector(FV).To address the problem of unbalanced distribution,we further propose several evaluation criteria.Comprehensive experiments are carried out on benchmark PETs2009 dataset.What we observe are,firstly,convolutional features outperform existing hand-crafted ones.Moreover,utilizing pre-trained deep convolutional features usually leads to better performance than models trained from scratch.Finally,simpler pre-trained models such as AlexNet turn out to generalize better than more complicated ones such as VGGNet.