DIGITAL TWIN-BASED DRIVING SIMULATION FOR AUTONOMOUS DRIVING IN DEVELOPING REGIONS

Abstract

The current developments in the field of trajectory prediction have led to the realization that precision in motion forecasting is important in complex and non-structured traffic setups. This paper describes a Transformer-based architecture, which builds spatial-temporal motion patterns directly on GPS trajectories with the help of self-attention and trajectory-relative feature encoding, making it capable of modeling longer-range dependencies in recent patterns than recurrent networks. The multi-head attention mechanism is an effective way to improve fine-grained motion understanding and, at the same time, is able to run in real time. The trained model is deployed into a SUMO-based digital twin, making it possible to keep prediction simulation in sync at 0.1 second intervals. GeoLife experimental results indicate good performance, with a validation loss of 0.000443, an RMSE of 0.021, an ADE of 0.025, and 97.27% accuracy at a threshold of 0.05 in centimeter accuracy and higher than LSTM based baselines.