Explainable LSTM-Based Cyclist Intention Prediction at Intersections

Increasing the safety of vulnerable road users (VRUs) in traffic has become a topic of general interest. Predicting cyclists’ turning intention in intersections can benefit safety applications in forecasting potential accidents. In this paper, we propose a bidirectional, stacked LSTM intention prediction model utilizing real-world smartphone cycling traces. We show that even imprecise GPS data are sufficient to predict right turns, and straight-going traces with a certainty of 90 % 45 m, and left turns 28 m before the intersection center, resulting in recognizing even the intention of the fastest cyclist in the data set 4.19 s before reaching the center. We further conduct an explainability analysis, including feature engineering, and SHapley Additive exPlanations (SHAP), highlighting the influence of GPS positions, and rotation vectors on our model. Lastly, we investigate the generalizability of our model on untrained intersections, showing first promising results for left turns of 90 % prediction probability 45 m before the intersection center, and probabilities of 90 % 20 m for straight-going traces, for an exemplary intersection.