Research On Crowd Counting And Density Estimation Based On Deep Convolutional Neural Network

The exponential growth of the world population and accelerated urbanization process have led to an increased number of activities such as sporting events,political rallies,public demonstrations etc.,thereby resulting in more frequent crowd gatherings in the recent years.Overcrowding will bring potential danger and easily lead to stampede.Therefore,crowd counting and density estimation become the hot topics in security field and has a wide range of applications such as video surveillance,traffic monitoring,public safety and urban planning.However,the task comes with many challenges such as scene variations in appearance,non-uniform distribution of people,high clutter,non-uniform illumination,occlusions,scale and perspective changes making the problem extremely difficult.On the basis of summarizing and analyzing existing methods,this paper designed and improved the crowd counting and density estimation models based on deep convolution neural network which were compared with other methods on the open dataset.The main work of this paper includes:(1)This paper fist expounded the significance of crowd counting and density estimation,and analyzed the detection-based and regression-based crowd counting methods.The methods based on convolution neural network were investigated and evaluated thoroughly.(2)To solve the problem of redundant structure,various parameters and difficult training of MCNN model,a single column deep convolution neural network model named Model-A was designed on the basis of its Small column network.The improved network consists of 6 convolution layers and 2 pool layers.Experiments proved that single column deep convolution neural network can also be robust to scale change of targets.(3)By further increasing the number of network layers,a deep convolution neural network based on feature fusion named Model-B was proposed.The deconvolution technique is used to upsample the feature maps.The feature maps of convolution layers with the accumulated downsample coefficient of 4 and 8 is concated.The network only use 3x3 convolution kernel with very small receptive field,and the high level neurons obtain larger receptive field by increasing the number of network layers.Experiments showed that the combination of shallow features and high-level features can extract richer features;(4)Model-C was proposed by improving the training data processing,loss function and training parameters of a more complex scale-adaptive convolution neural network(SaCNN).Inspired by DenseNet,the method of dense connection was used to strengthen the features reuse and propagation in the shallow network of Model-C,and the scale adaptive convolution neural network model with dense connection name Dense_Model-C was designed and implemented.(5)Experiments are carried out on a challenging ShanghaiTech dataset using the mean absolute error(MAE)and mean square error(MSE)for evaluating.The geometric adaptive Gauss kernel is used to generate real density maps.In the training process,random croping is applied to speed up and increase the diversity of samples.The results showed that the Model-A is better than MCNN and much better than that single column networks in MCNN.The MAE and MSE of Model-B is lower than Model-A,which are competitive to the state of the art methods.The results of Model-C are better than that of Model-A and Model-B,which outperform the state of the art methods.The MAE and MSE of Dense_Model-C is slightly higher than Model-C,but it reduces the model parameters and can be convergence faster.