A Novel Crowd Counting Method via Deep Convolutional Neural Network

With the rapid growth of population and urbanization, urban management departments pay more and more attention to the monitoring of dense crowd. So it is important to estimate the number of people accurately for scene understanding. In this paper, we propose a new crowd counting framework with deep convolutional neural network. In order to overcome the perspective problem of the camera, the proposed framework uses multi filters with different sizes to enhance the discriminative power of the framework. In the proposed work, end-to-end approach is utilized to estimate the crowd density map. In order to speed up the training of the model and reduce the number of parameters, we processed the data with 1×1 filters, which could reduce the numbers of channels. Finally, the proposed framework combines the feature layers calculated by different sizes of filters to output the density map. In this way, we allow the model to merge different features calculated by different sizes of filters, rather than setting them manually. To evaluate the performance of the algorithm, the proposed model is compared with seven popular algorithms on three open datasets (ShanghaiTech dataset, UCF_CC_50 dataset and WorldExpo'10 dataset). The experimental results show that the proposed algorithm is effective. Specifically, comparing with the existing best-performing methods, we achieve an improvement of 7% the ShanghaiTech dataset and 8% on the most challenging dataset UCF _CC_50 dataset.