PROSPECT aims at developing a new generation of proactive safety systems to protect Vulnerable Road Users (VRUs), with an emphasis on pedestrians and cyclists. To improve sensor effectiveness, PROSPECT will expand the scope of scenarios addressed by sensors already on the market, enhancing their overall performance.
Interactions between vehicles and VRUs were investigated in real traffic situations to better understand critical situations and identify factors that lead to conflicts. As a result, VRU and vehicle modelling will be more effective, allowing safety systems to react earlier, without increasing false activation rates. Accident studies highlighted the most relevant use cases, and further naturalistic observations provided information that could not be inferred from accident databases regarding these use cases, such as trajectories and kinematic data (speed, acceleration, TTC or PET) throughout the conflict evolution. Data was also collected on VRU’s behaviors which forecast their intent in the near future (i.e. positional data, gestures). Lastly, naturalistic observations were used to look for correctly managed situations by the road users that could lead to false alarms in existing sensors.
Two kinds of naturalistic observations were undertaken in three countries. A first data set (France and Hungary) was collected from on-site observations by infrastructure-mounted cameras. A second data set was collected by cars equipped with sensors and cameras (Hungary and Spain) to observe interactions with surrounding VRUs.
Only situations of conflict with close proximity between road users both in space and time were studied. This important criterion qualified an encounter as a conflict. Low speed conflicts were excluded. Several hundred conflicts were collected, each classified according to use cases and annotated using a common grid. Different categories of parameters were investigated to describe: environmental conditions (light, precipitation, road surface, traffic density, etc.), infrastructure (layout, dedicated lanes, speed limit, etc.), VRU characteristics (type, equipment, etc.), encounter (visibility, right of way, yielding, conflict management, estimated impact point, etc.), intent (head/torso orientation, gesture, flashing indicator), kinematics and trajectories.
Start and end timestamps were recorded for time dependent parameters such as yielding, head movements, etc.
Finally, variants of use cases were obtained to describe potential conflict evolutions and determinant factors of this evolution.
As annotations of conflicts were based on subjective evaluation of observers, training was required. Although training sessions were organized, materials differed between observations which could lead to some distortion. However, including objective data such as kinematics and trajectories mitigated data validity concerns. Severity of conflicts, for example, was first assessed by subjective measure (as filtering process), then revised by taking into account kinematic data as a more objective measure. We also considered inconsistent accuracy level of video processing algorithms for spatial data (trajectories and kinematics).