Simulating Driver-Pedestrian Interactions and Intentions Inference

The appearance of autonomous vehicles (AVs) has become a crucial and ever more emerging topic, giving way to new discoveries in the behavioral aspect of robotics. It is imperative that AVs know how to interact with pedestrians during driving contexts. Interaction between drivers (either active or passive ones) and pedestrians is something crucial to driving experiences in intersections, signaled or non-signaled crossings. If these situations are not well controlled and interaction outcomes are not predicted, risky situations occur. Thus, the creation of models pertaining to such crossing scenarios is a powerful tool to infer over real-life ones, where testing hypotheses may be too dangerous. When creating these models one needs to take into consideration the importance of extrinsic and intrinsic factors of driver-pedestrian interactions. These latter ones, ie. the communication between the two, can give drivers useful hints of intentions by pedestrians. Knowing such intentions means knowing how to react to pedestrian crossing scenarios. Studies found that non-verbal communication is very useful in asserting confidence by pedestrians. Related work is extensive. Many authors have already used factors like vehicle speed, among many other factors to build predictive models that can output confidence of the pedestrian crossing. Some authors used neural network approaches to estimate pedestrian poses in crossing scenarios. If AVs need to predict such intentions in driving contexts, predictive models that take into consideration this communication must be built. Thus, the contributions that this dissertation brings are twofold. Firstly, the study of the relevance of non-verbal communication when predicting outcomes of crossing scenarios. It is important to justify the need or lack thereof of computing non-verbal communication and using it along other factors in predictive models. Secondly, the creation of a methodology to better study which factors influence driver-pedestrian interactions, and how information about them can be extracted directly from the source into results. These results can be used for predictive models to be employed in simulations. The solution was elaborated using a dataset created from a set of experiments and some image segmentation techniques. These experiments led subjects to drive around in a virtual cockpit in an environment filled with pedestrians. Their actions during pedestrian encounters permitted the analysis and extraction of conclusions. Obtained results emphasize the need of knowing how to react to non-verbal communication alongside the normally used metrics to predict pedestrian crossings. They relate the main factors of these driver-pedestrian interactions and provide better knowledge about their influence in the outcomes of such scenarios. This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement N. 723386 (SIMUSAFE)