Enhanced and Pruned Motion Planning Based on Bird’s-Eye View

To address the imbalance in task allocation between the encoder and decoder in end-to-end autonomous driving systems, we propose a method based on the semantic bird's-eye view and enhanced and pruned motion planning. This study utilizes a depth estimation network to infer pixel depth and combines camera intrinsic and extrinsic parameters to map image features to bird's-eye view features. Subsequently, an enhancement module and a pruning module are introduced to extract key information. Additionally, a feature evaluation module connects the two modules, allowing the pruning module to learn from the features extracted by the enhancement module. Finally, a GRU planning module is employed to predict the vehicle's future trajectory. An evaluation of the NoCrash benchmark in the Carla town2 simulation platform demonstrated that the proposed method achieved an average driving score of 84.25 under known weather conditions, validating its effectiveness.