Radar and Vision Deep Multi-Level Feature Fusion Based on Deep Learning

Based on deep learning, a perception network structure of radar and vision depth multilevel feature fusion is designed to realize the effective perception of intelligent driving in foggy scenes. Aiming at the problem of low efficiency when traditional feature fusion methods are applied to heterogeneous sensors with asymmetric feature information, spatial attention fusion is used to realize asymmetric feature fusion of radar and vision. Aiming at the problem that the fusion level of radar and vision low-level single features is shallow, and the fusion features are difficult to transfer to the deep network, a deep multilevel feature fusion structure is constructed to realize the multi-stage fusion of radar and vision features. At the same time, the attention mechanism module is introduced to integrate the fusion features and extract key information. Finally, the validity of the network is verified by testing. The obtained defogging results are sent to the network for testing, which further verifies the effectiveness of the fusion defogging algorithm. Through comparative experiments, it is proved that the fusion perception network can effectively improve the detection accuracy and recall rate of the intelligent vehicle perception system in foggy scenes.