Scientists at Honda Research Institute in the United States have compiled LOKI, a dataset that can be used to train models that predict the paths of pedestrians and vehicles.
The authors of the new work hypothesized that the models are best trained by predicting the short-term and long-term goals of objects around. The resulting model effectively plans the movements of the robot or vehicle based on the predicted movements.
The researchers plan to develop an architecture that considers both short-term and long-term goals – these are the main components in assessing the intentions of a pedestrian or a car.
For example, there is a car that wants to turn left at an intersection. It is important to take into account not only the dynamics of transport, but also how the intention can change depending on various factors: due to one’s own desire, other road users or obstacles.
Next, the algorithm first codes the past experience from which the model was trained to predict what the long-term and short-term consequences might be.
The model assigns labels to all surrounding objects – these are “intentions” that can change as you move, “the environment”, for example, road signs and trees that affect the intentions of agents, as well as “contextual labels” such as weather and road conditions.
The researchers evaluated their model in a series of tests and found it outperforms other modern trajectory prediction methods by 27%.
The developers believe that the model can be used to improve the safety and performance of autonomous vehicles. In addition, other research groups can use the LOKI dataset to prepare their own models to predict the paths of pedestrians and vehicles on the road.