Predicting Bicyclist Maneuvers using Explicit and Implicit Communication

Abstract Explicit and implicit communication cues are part of the communication strategies often employed by bicyclists in traffic to reveal their intentions to other road users. For human drivers, such types of information may be understandable in a specific social and cultural context. Present automated driving functions mostly rely on vulnerable road user (VRU) dynamic and operational features to perform short-term predictions of nearby VRU behaviour. The understanding of bicyclist implicit and explicit communication behavior can provide an extra layer of information in predicting their movement in traffic, confine the solution space to specific maneuver archetypes and expand the short-term time window for the bicyclist trajectory prediction. This paper focuses on the problem of predicting the bicyclist maneuver at an intersection approach using exclusively bicyclist explicit and implicit communication cues. Data is first evaluated using random forests for feature selection to determine the most important features for the prediction task. The next step involves training and testing different classifiers such as k-nearest neighbors, extra trees, decision trees, random forests, logistic regression, linear support vector machines, and the gaussian naive-bayes classifier to define the most efficient methodology to address the maneuver classification problem. Results suggest that the gaussian naive-bayes classifier provides the best performance on the test data with a 69.4% accuracy score during training and 67.0% accuracy score during testing. The proposed method has the potential to support and cooperate with existing solutions for bicyclist trajectory prediction as part of automated driving functions, effectively increasing traffic safety.