Concise Convolutional Neural Network for Crowd Counting

Utilizing convolutional neural network (CNN) for estimating the crowd count in a still image has taken some progress. But the existing models or algorithms are too complex for actual applications and their real-time performance have not been effectively verified. In this paper, we propose a concise and effective CNN model with only five convolutional layers. Our proposed model allows the input image to be of any size or resolution. The model maps the input image to its crowd density map. By integrating the density map we can get the crowd count. The true density map is computed based on geometry-adaptive kernels which can alleviate the perspective problems. We report the performance in terms of mean absolute error, which is a measure of accuracy of the method. We conduct extensive experiments on major crowd counting datasets to verify the effectiveness of the proposed model and apply it to the actual situation successfully. In addition, we created a new dataset to verify the transfer learning performance and real-time performance of our model. Experiments show that it has great transfer learning performance and real-time performance.