A High-robustness and Low Resource-consumption Crowd Counting Model

Abstract

A major problem of counting high-density crowded scenes is the lack of flexibility and robustness exhibited by existing methods, and almost all recent state-of-the-art methods only show good performance in estimation errors and density map quality for select datasets. The biggest challenge faced by these methods is the analysis of similar features between the crowd and background, as well as overlaps between individuals. Hence, we propose a light and easy-to-train network for congestion cognition based on dilated convolution, which can exponentially enlarge the receptive field, preserve original resolution, and generate a high-quality density map. With the dilated convolutional layers, the counting accuracy can be enhanced as the feature map keeps its original resolution. By removing fully-connected layers, the network architecture becomes more concise, thereby reducing resource consumption significantly. The flexibility and robustness improvements of the proposed network compared to previous methods were validated using the variance of data size and different overlap levels of existing open source datasets. Experimental results showed that the proposed network is suitable for transfer learning on different datasets and enhances crowd counting in highly congested scenes. Therefore, the network is expected to have broader applications, for example in Internet of Things and portable devices.