Intention Analysis for Vulnerable Road Users

Kurzfassung

In the last few years, the research on the protection of Vulnerable Road Users (VRU) has been receiving increasing attention. One of the main stakes of this research is to develop systems able to predict the intention of a VRU and more specifically its future trajectory. In order to make such predictions, multiple solutions have been studied: solutions relying on Gaussian Processes, on Long Short Term Memory (LSTM) neural networks, on encoders and decoders, on Convolutional Neural Network (CNN) or on Generative Adversarial Network. All those solutions are designed to model the relation between the pedestrian’s past trajectory and its future trajectory in order to predict accurately the pedestrian’s intention. Such forecasting abilities would allow to warn pedestrians of an incoming danger on their trajectory an thus avoid possible collisions. This knowledge could also be applied for Advanced Driver Assistance Systems (ADAS) and autonomous vehicles to help anticipating nearby pedestrians’ behaviours. However, most of the previously mentioned models have only been tested on short-term predictions (less than 5 seconds) whereas longer-term predictions would be necessary to warn early enough and avoid an upcoming collision. Moreover, the models ability to forecast trajectories close from the true trajectories have been tested, but verifying whether those models are able to discriminate between dangerous and safe trajectories in given situations is less studied. Thus, the objectives are to evaluate several model architectures on longer-term predictions and in predefined scenarios. The term 'Vulnerable Road Users' often refers to pedestrians, cyclists and motorcyclists. However, according to Hollaender et al., it is important to differentiate between the multiple road users. Thus, in this work, only pedestrians have been considered but the models and scenarios could be adapted to other types of VRU.