SocialNet: Detecting Social Distancing Violations in Crowd Scene on IoT devices

Recently, the COVID-19 pandemic has affected the world and spread in the majority of the countries. To decrease the number of infections, experts suggested people practice social distancing by maintaining a distance of six feet apart. It is hard to monitor this restriction by only a traditional surveillance system. Existing methods used deep learning to tackle this problem by designing a Deep Convolutional Neural Network (DCNN). However, these methods do not accommodate for low-power systems such as Internet-of-Things-based devices. In this paper, we propose SocialNet, a novel network design that can detect violations of social distancing in public crowd scene. SocialNet is composed of two components, (1) The detector backbone and (2) The Autoencoder. In the detector backbone, the network generates the bounding boxes of the human/person category. In the Autoencoder, the network learns to predict a score which represents the violation between each bounding box and the others. The input image is divided into nine patches and each patch goes through the Autoencoder to predict the violation score. The Autoencoder consists of the encoder which produces the latent vector for the decoder network. The decoder network outputs a real-valued vector of length fifteen as scores for each patch in the image. Furthermore, SocialNet uses mixed precision which makes it suitable for low-power devices.