Design of crowd counting system based on improved CSRNet

Abstract

An advanced crowd counting algorithm based on CSRNet has been proposed in this study to improve the long training and convergence times. In this regard, three points were changed from the original CSRNet: (i) The first 12 layers in VGG19 were adopted in the front-end to enhance the capacity of the extracting features. (ii) The dilated convolutional network in the back-end was replaced with the standard convolutional network to speed up the processing time. (iii) Dense connection was applied in the back-end to reuse the output of the convolutional layer and achieve faster convergence. ShanghaiTech dataset was used to verify the improved CSRNet. In the case of high-density images, the accuracy was observed to be very close to the original CSRNet. Moreover, the average training time per sample was three times faster and average testing time per image was six times faster. In the case of low-density images, the accuracy was not close to that of the original CSRNet. However, the training time was 10 times faster and the testing time was six times faster. However, by dividing the image, the count number came close to the real count. The experimental results obtained from this study show that the improved CSRNet performs well. Although it is slightly less accurate than the original CSRNet, its processing time is much faster since it does not use dilated convolution. This indicates that it is more suitable for the actual needs of real-time detection. A system with improved CSRNet for counting people in real time has also been designed in this study.