Autonomous driving, bird’s eye view, multiscale feature fusion, LiDAR, feature alignment, Electrical engineering. Electronics. Nuclear engineering, TK1-9971

Scene understanding and multisource data fusion are critical challenges in autonomous self-driving systems.In particular, optimizing information fusion strategies for three-dimensional Bird’s Eye View (BEV) scene recognition tasks is crucial for accurate perception and decision-making in dynamic environments. This study proposes a novel architecture that integrates multiscale feature extraction and crossmodal structural alignment to enhance the representation and detection capabilities of BEV features. Specifically, we employ a DCN-based block for visual feature extraction, comprising layer normalization (LN), feedforward networks (FFNs), and the Gaussian Error Linear Unit (GELU) activation function, aligned with the Vision Transformer (ViT) paradigm to improve feature modeling. To fully utilize multiscale information, a dedicated multiscale feature fusion block is introduced to extract expressive scene features within the feature space. Furthermore, we leverage LiDAR to generate LIDAR BEV features and propose a feature alignment block to enhance the complementarity between camera and LiDAR BEV features. The proposed architecture effectively supports precise scene recognition and adaptive decision-making in multi-sensor fusion environments, providing robust perception capabilities for autonomous driving in complex scenarios.