Deeper multi-column dilated convolutional network for congested crowd understanding

In highly congested crowd scenes, it is hard to generate high-quality density maps because the crowd and background are highly mixed such that it is difficult to distinguish them. To alleviate the issue, this paper presents a deeper multi-column dilated convolutional network (DMDCNet) method, which is capable of extracting sufficient semantic features for crowd understanding in highly congested crowd scenes. In DMDCNet, there are two modules: feature extractor and density map estimator. Feature extractor is a VGG-16-based convolutional neural network (CNN), which is able to extract low-level features contained in crowd images. Density map estimator is designed as a multi-column structure of dilated convolutional neural networks (DCNNs) to further extract the deeper information and capture multi-scale contextual information, which could generate high-quality density maps from the input images. Furthermore, multi-column DCNNs in DMDCNet can effectively alleviate the “gridding” problem caused by the dilated convolution framework. Extensive experiments on several commonly used benchmark datasets are conducted to demonstrate the proposed DMDCNet, which shows that DMDCNet is comparable with the recent state-of-the-art methods.