Machine Learning based Motion Forecast for Automotive Applications
Buch in deiner Nähe kaufen
Beschreibung
The prediction of the future motion of road users is a critical task in supporting advanced driver-assistance systems (ADAS). It is of even more importance for self-driving vehicles in enabling the planning and execution of safe driving maneuvers. It is a challenging task due to the complexity of the possible driving scenarios. The intention of other road users is not directly observable, and often multiple future trajectories can be plausible, for example, turning right or continuing straight at an intersection. In addition, there are numerous possible interactions among road users and the driving environments, which are practically not feasible for explicit modeling. In this thesis, these challenges are tackled by jointly learning and predicting the motion of all road users in a scene, using a novel convolutional neural network (CNN) and recurrent neural network (RNN) based architecture, called CASPNet. Moreover, by leveraging grid-based input and output data structures, the computational cost is decoupled from the number of road users, and multi-modal predictions become inherent properties of our proposed method. Evaluations on public datasets show that the proposed approach reaches state-of-the-art performance, reaching first place in the nuScenes prediction benchmark. Pedestrians are especially vulnerable in traffic scenarios. This thesis introduces a method that detects and integrates pedestrian body and head orientation into a prediction system to predict the movements of pedestrians more accurately. Finally, CASPNet output modality was extended to predict the future ego path, integrated into an ADAS system, and successfully validated in real-world closed-loop operation.
Autor*in
Maximilian Schäfer
Themen in »Machine Learning based Motion Forecast for Automotive Applications«
Motion Forecast Trajectory Prediction Deep Learning Artificial Intelligence