Self-Attention based encoder-Decoder for multistep human density prediction

Multistep Human Density Prediction (MHDP) is an emerging challenge in urban mobility with lots of applications in several domains such as Smart Cities, Edge Computing and Epidemiology Modeling. The basic goal is to estimate the density of people gathered in a set of urban Regions of Interests (ROIs) or Points of Interests (POIs) in a forecast horizon of different granularities. Accordingly, this paper aims to contribute and go beyond the existing literature on human density prediction by proposing an innovative time series Deep Learning (DL) model and a geospatial feature preprocessing technique. Specifically, our research aim is to develop a highly-accurate MHDP model leveraging jointly the temporal and spatial components of mobility data. In the beginning, we compare 29 baseline and state-of-the-art methods grouped into six categories and we find that the statistical time series and Deep Learning Encoders-Decoders (ED) that we propose are highly accurate outperforming the other models based on a real and a synthetic mobility dataset. Our model achieves an average of 28.88 Mean Absolute Error (MAE) and 87.58 Root Mean Squared Error (RMSE) with 200,000 pedestrians per day distributed in multiple regions of interest in a 30 minutes time-window at different granularities. In addition, the geospatial feature transformation increases 4% further the RMSE of the proposed model compared to the state of the art solutions. Hence, this work provides an efficient and at the same time general applicable MHDP model that can benefit the planning and decision-making of many major urban mobility applications.