ABSTRACT :

Crowd counting is an important research topic in the field of computer vision. The multi-column convolution neural network (MCNN) has been used in this field and achieved competitive performance. However, when the crowd distribution is uneven, the accuracy of crowd counting based on the MCNN still needs to be improved. In order to adapt to uneven crowd distributions, crowd global density feature is taken into account in this paper. The global density features are extracted and added to the MCNN through the cascaded learning method. Because some detailed features during the down-sampling process will be lost in the MCNN and it will affect the accuracy of the density map, an improved MCNN structure is proposed. In this paper, the max pooling is replaced by maxave pooling to keep more detailed features and the deconvolutional layers are added to restore the lost details in the down-sampling process. The experimental results in the UCF_CC_50 dataset and the ShanghaiTech dataset show that the proposed method has higher accuracy and stability.

EXISTING SYSTEM :

Crowd counting is used to calculate the total number of people in images or video frames. The crowd counting methods can be divided into three categories: the direct counting method based on target detection, the indirect method based on feature regression and crowd counting based on deep learning. In the relevant researches based on target detection [1]–[5], Lin et al. [1] proposed to use Haar wavelet transform to extract the feature area of the head-like contour and build the SVM classifier to classify the feature area. Gardzinski et al. [2] proposed to use shape contour of body to achieve crowd detection and crowd density estimation. All of these methods are suitable for the scenes with low density crowd, but the detection accuracy will decrease in the case of high density crowd. In the relevant researches based on feature regression [6]–[10], the regression relationships between image features and the number of people are established for crowd counting. Chan and Vasconcelos [7] proposed to use low-level features and Bayesian regression to improve the robustness and adaptability of the regression model.

PROPOSED SYSTEM :

with the rapid development of deep learning and big data [11]–[14], crowd counting methods based on deep learning are proposed gradually. Zhang et al. [15] proposed a cross-scene crowd counting model. It was trained alternately through two learning objectives, density map and global number. This algorithm is implemented based on single-column CNN. However, it is not suitable for the change in the scale of crowd. Zhang et al. [16] proposed to use the MCNN with three branch networks for crowd counting. Different receptive fields were used in each branch network, and this improved MCNN could adapt to the change in the scale of the crowd. They also introduced a new dataset ShanghaiTech for crowd counting. Boominathan et al. [17] proposed to combine the features of shallow and deep convolutional neural networks to improve spatial resolution. Sindagi and Patel [18] proposed a multi-task network which combined the high-level prior with the density estimation. Sam et al. [19] proposed Switch-CNN for crowd counting. In this network, a classifier was trained and an appropriate regressor was selected for input patches. Shi et al. [20] proposed to aggregate multiscale features into a compact single vector and used deep supervised strategy to provide additional supervision signal. Fu et al. [21] proposed to use the LSTM structure to extract the contextual information of crowd region. Liu et al. [22] proposed to add an attention module to adaptively select the counting mode used for different positions on the image. Yang et al. [23] proposed to use the MMCNN for robust crowd counting. In this work, the location, detailed information and scale variation were taken into account to generate density map in order to improve the robustness of crowd counting method. Generally, these algorithms have good performances in the crowd counting, but the performances of these methods were not effective when the crowd distribution is uneven [24], [25]. In order to solve the problem of inaccurate counting caused by uneven crowd distribution, the global density feature is extracted and used in this paper. A convolutional neural network with global density feature by using multi-task network cascades (MNCs) [18], [26] is proposed. In order to generate a more comprehensive density map, the maxave pooling layers are used to keep more features of the image. Meantime, the deconvolutional layers are added to the convolutional neural network in order to restore the lost details in down-sampling process. It will help to improve the accuracy of density map and further improve the accuracy of crowd counting.

SYSTEM REQUIREMENTS
SOFTWARE REQUIREMENTS:
• Programming Language : Python
• Font End Technologies : TKInter/Web(HTML,CSS,JS)
• IDE : Jupyter/Spyder/VS Code
• Operating System : Windows 08/10

HARDWARE REQUIREMENTS:

 Processor : Core I3
 RAM Capacity : 2 GB
 Hard Disk : 250 GB
 Monitor : 15″ Color
 Mouse : 2 or 3 Button Mouse
 Key Board : Windows 08/10