Prediction of Human Intention in Vehicles, Pedestrians and Bicyclists Interactions from Videos

Abstract

Humans' intention prediction in vehicles, pedestrians and bicyclists interactions can help autonomous vehicles and human drivers to plan their routes in a safer manner and optimise the use of roads space. Several studies have tried to estimate humans' intention when interacting with other agents at crossroads using hand-craft features, motif analysing, and machine learning approaches. However, many of them are not accurate enough due to insufficient consideration of the surrounding agents and limited observations (occlusions, inaccurate pose and location estimation) caused by camera angles. This study proposed a multi-branch GRU encoder-decoder (MBDED) model to predict pedestrians' and bicyclists' intention when contenting with vehicles at intersections by analysing the properties of directly and indirectly involved road agents. The model was constructed with an encoder-decoder structure using four GRU branches, where each encodes a set of information of mobile agents. The network was trained, validated, and tested on unsignalised and uncontrolled intersections. The system predicts the vulnerable road users' intentions with 96% accuracy, 91% precision, and 93% recall at 2 seconds before the intersections happen, which provide reliable reference for autonomous vehicles navigation and sufficient time for human driver to react.