This paper presents the results of an empirical study on a vision-based trajectory prediction approach for usage in urban environments based on conditional imitation learning with focus on bicycles. The approach uses neural networks to predict the changes in heading and speed of an ego vehicle for a specified prediction horizon and sample time which can be shared over V2X for collision avoidance or used in further advanced driver assistance systems. Predicting changes allows a location-independent trajectory representation based on the vehicle's current speed and heading. The prediction is carried out for the four predefined maneuvers “turn left”, “turn right”, “go straight” and “lane follow”. Several neural networks with different input data configurations are evaluated for the same prediction task. The available network inputs are images from a front-facing camera, heading information, vehicle speed and turning indicator signals. For training the networks a dataset was created from several hours of riding an electric bicycle in structured and unstructured environments of a city. In addition the networks have been implemented on an embedded GPU platform to investigate the real-time capability of the presented approach. The networks were configured and trained to predict five seconds into the future with a sample time of 0.2 s. The results of each network are presented and compared against each other. The results show that the implemented networks are capable of predicting a trajectory in short to mid-term (1 - 3 s) with good accuracy while having difficulties predicting turn maneuvers in long-term range (3 - 5 s) due to the lack of road layout information.
An Empirical Study on Ego Vehicle Trajectory Prediction for Bicycles in Urban Environments Based on Conditional Imitation Learning
19.09.2021
1142319 byte
Aufsatz (Konferenz)
Elektronische Ressource
Englisch
Full Vehicle Trajectory Planning Model for Urban Traffic Control Based on Imitation Learning
Transportation Research Record | 2022
|Transportation Research Record | 2014
|Vehicle Trajectory Prediction Based on GAT and LSTM Networks in Urban Environments
DOAJ | 2024
|