Cyclist trajectory prediction using bidirectional recurrent neural networks

Predicting a long-term horizon of vulnerable road users’ trajectories such as cyclists become an inevitable task for a reliable operation of highly and fully automated vehicles. In the literature, this problem is often tackled using linear dynamics-based approaches based on recursive Bayesian filters. These approaches are usually challenged when it comes to predicting long-term horizon of trajectories (more than 1 sec). Additionally, they also have difficulties in predicting non-linear motions such as maneuvers done by cyclists in traffic environments. In this work, we are proposing two novel models based on deep stacked recurrent neural networks for the task of cyclists trajectories prediction to overcome some of the aforementioned challenges. Our proposed predictive models have achieved robust prediction results when evaluated on a real-life cyclist trajectories dataset collected using vehicle-based sensors in the urban traffic environment. Furthermore, our proposed models have outperformed other traditional approaches with an improvement of more than 50% in mean error score averaged over all the predicted cyclists’ trajectories.