Prediction

Prediction is a critical component of autonomous driving and directly affects system safety, reliability, and user experience. In complex and changing traffic environments, accurate predictions help autonomous vehicles make decisions and plan ahead, reduce unnecessary emergency braking or evasive maneuvers, and avoid potential collision risks, improving both safety and passenger experience.

Typically, prediction tasks take road users’ recent state information (e.g., positions, velocities, and headings over 1s or 3s) and map information (e.g., lanes, road markings, and traffic lights) as input, and output their future states over a horizon such as 3s or 5s (trajectory coordinates or distributions).

The prediction problem is broken down into:Motor-Vehicle Trajectory Prediction Task、Vulnerable road user (VRU) trajectory prediction task、Multi-target trajectory prediction task。

Motor-Vehicle Trajectory Prediction Task

Motor vehicle trajectory prediction refers to predicting the future trajectories of other motor vehicles on the road. The difficulty of this task is that the behavior of motor vehicles is affected by many factors such as the driver's operating intentions, traffic rules, road conditions, and the surrounding environment. For example, a vehicle may suddenly change lanes, slow down, or accelerate. These behaviors are often sudden and may be affected by the dynamics of the vehicle ahead. In addition, different driving habits and complexity of road conditions will also lead to diversity of vehicle behavior, making prediction more difficult.

VRU Trajectory Prediction

VRU trajectory prediction mainly refers to predicting the future movement trajectories of non-motorized vehicles such as bicycles and electric scooters, as well as pedestrians. The challenge is that non-motor vehicles are extremely flexible. Non-motor vehicles are usually not restricted by fixed lanes. The speed and direction of non-motor vehicles change frequently, and their driving routes may be more arbitrary. Drivers' behavior patterns are difficult to predict, and some drivers may lack professional driving skills, resulting in higher uncertainty and unpredictability of behavior. Pedestrian behavior is highly uncertain and diverse. Pedestrians may suddenly change direction, stop, or cross roads, and their behavior is affected by personal purposes, traffic signals, and other factors.

Multi-Agent Prediction

Multi-objective trajectory prediction refers to simultaneously predicting the movement trajectories of multiple different types of road participants (including motor vehicles, non-motor vehicles and pedestrians) in the future in an autonomous driving scenario. The difficulty of this task is that the behavior patterns and movement patterns of different objects vary significantly and are highly dependent on interactions with each other and changes in the surrounding environment. Multi-objective trajectory prediction must also take into account the interactions between different objects. For example, a vehicle may slow down and stop in front of a zebra crossing to give way to pedestrians; a motor vehicle may adjust its speed or change lanes to avoid non-motor vehicles. These complex interaction patterns require that predictive models not only have powerful data processing and learning capabilities, but also need to be able to understand and simulate human behavioral logic and social norms.