Prediction is the estimation of how the future will evolve. It can be as diverse as a human’s motion, the weather or the next action performed by a human co-worker. While most of us have experienced the uncertainty in predicting the weather, the later example is qualitatively different. Humans have the freedom to choose their next step or behavior and it depends on preferences, experiences or context. In a similar way, machines are trained to decide based on a target function.
In a system, intentions of all participants – human or machine, decision maker or bystander – have to be predicted correctly.
At HRI-EU, we decompose the prediction task into a variety of different sequences, each following a certain qualitative pattern of assumptions and constraints. These sequences can merge and diverge depending on the continuous perception and new information of the environment. If a decision is required, sequences are combined which take into account their different levels of reliability. An illustrative example is the combination of context-based and physical prediction outlined on the previous page.
Prediction of Highway Lane Changes
Predicting the lane change intentions of other traffic participants is an important challenge for autonomous driving and driver assistance systems. In this research, a physical prediction is combined with a context-based prediction and the ego-vehicle‘s behavior influence.
The physical prediction uses observations of recent lateral vehicle positions and compares them to models of typical lane change trajectories. The resulting prediction is highly accurate, but only capable of predicting behaviors once they started.
The context-based prediction is based on cognitive models comprising the influence of the driving context on the lane change intention. It evaluates relations between the target and its context vehicles to estimate a future lane change. For example, a driver tends to change lane when approaching a slower vehicle with a sufficient gap on the next lane. This provides a lane change prediction before any physical movement can be observed.
Given a probability for a behavior of the ego-vehicle, the influence of its decision on the prediction is inferred. The combined method allows to make foresighted behavior decisions that robustly factor in future changes in the driving situation.
For more information
T. Weisswange, et al., “intelligent Traffic Flow Assist: Optimized Highway Driving Using Conditional Behavior Prediction”, IEEE Intelligent Transportation Systems Magazine, 2019