End-to-end Network for Autonomous Vehicles Based on a Bird’s Eye View Perception Module

End-to-end autonomous driving reduces errors and information loss during inter-module communication by directly predicting control commands from input sensor data. The trajectory-guided control prediction (TCP) method fuses features from images, vehicle states, and target points, and inputs them to a gated recurrent unit (GRU)-based action predictor module, which outputs steering and acceleration commands. However, data imbalance during training makes it difficult to extract meaningful features in data-sparse scenarios, leading to degraded driving performance. Therefore, we propose a Bird’s Eye View perception module (BPM) to enable robust driving even in data-sparse situations. Additionally, BPM extracts spatial and temporal features, and the GRU-based action predictor module leverages these features to predict trajectories and control commands. We trained and evaluated our method and TCP using the MORAI simulator. The training data included 75.84% lane following, 15.13% left/right turns, and 9.03% avoidance scenarios. In data-sparse scenarios (e.g., left/right turns, avoidance), our method reduced the lateral error by 0.83 m and the collision rate by 27.78%.