Improving Efficiency and Generalisability of Motion Predictions With Deep Multi-Agent Learning and Multi-Head Attention
Djamel Eddine Benrachou, Sébastien Glaser, Mohammed Elhenawy, Andry Rakotonirainy
Abstract
Automated Vehicles (AVs) have been receiving increasing attention as a potential highly mechanised, intelligent, self-regulating futuristic mode of transport. AVs are predicted to address limitations and human factors associated with traditional modes of transportation. Beyond the typical operations of AVs which can perform rudimentary tasks, the intelligent embedded program fit in to process challenging scenarios and deep multi-dimensional/ agent intents and interaction of the roadway is the grey area yet to be explored to design an exclusive encoding of social functionality and operation in order to address human factors causing road crashes. The aim of this study is to design a data-driven prediction framework for AVs that utilises multiple inputs to prove a multimodal, probabilistic estimate of the future intentions and trajectories of surrounding vehicles in freeway operation. Our proposed framework is a deep multi-agent learning-based system designed to effectively capture social interactions between vehicles without relying on map information. Our approach excels in capturing the high-level behaviours of multiple vehicles and generating a multi-modal trajectory forecast. It employs a multi-headed neural architecture to learn from social interactions between vehicle pairs and generates diverse trajectories proportional to predicted target intents, thus enabling feature fusion. Additionally, a multi-head self-attention mechanism is incorporated for prediction refinement. We achieved a good prediction performance with a lower prediction error in real traffic data at highways. Evaluation of the proposed framework using the NGSIM (US-101 and I-80) and HighD datasets shows satisfactory prediction performance for long-term trajectory prediction of multiple surrounding vehicles. Additionally, the proposed framework has higher prediction accuracy and generalisability than state-of-the-art approaches.