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