Cyclist Intent Prediction using 3D LIDAR Sensors for Fully Automated Vehicles

Abstract

One of the main barriers against the full deployment of autonomous vehicles in urban traffic environments is the understanding of the intentions and behaviours of the human around them. Moreover, understanding and predicting intentions of vulnerable road users such as cyclists is still one of the most challenging tasks. In this work, we are proposing a novel framework for the task of intent prediction of cyclists via hand signalling from point cloud scans. We utilised our developed data generation pipeline for generating synthetic point cloud scans of cyclists doing a set of hand signals in urban traffic environments. Then, we feed a sequence of the generated point cloud scans to our framework which jointly segments all cyclists instances and predicts their most probable intended actions in an end-to-end fashion. Our proposed framework has achieved superior results with 83% in F1-Measure score over the testing split of our generated dataset. Additionally, the proposed framework outperformed other compared baseline approaches with more than 39% improvement in F1-Measure score.