Pre-Training Convolution Network for Crowd Counting

Image crowd density estimation is widely used in video surveillance, traffic surveillance and public safety. Recently, the convolutional neural network based approach has shown better results in crowd counting than the traditional approach. However, there are still a lot of limitations and difficulties in practical application :large-range number of people, changeable environment result in the current methods fail work well. In addition, Due to the lack of practical data, many methods suffer from over-fitting in varying degrees. To solve these two problems, first, we built a data collector and annotator, which can generate a large number of synthetic crowd scenes without any labor cost, and annotate them automatically. Secondly, we use pre-training method to pre-train on the synthetic dataset, and then fine-tune with the real data to effectively improve the performance of the model in the real scene. Extensive experiments show that the method is achieves the state-of-the-art performance on four real datasets.