Radar Fusion Monocular Depth Estimation Based on Dual Attention

In this article, we explore the integration of multimodal data into monocular depth estimation. Monocular depth estimation is performed by fusing RGB data with sparse radar data. Since the existing fusion method does not take into account the correlation between the two types of data in the channel and in space, it lacks the representation of the global information relationship on the channel and in space. Therefore, we propose a feature fusion module (DAF) based on the dual attention mechanism. The dual attention fusion module improves the global information representation capability of the model by modeling the dynamic and non-linear relationship of the two kinds of data in the channel and space, adaptively recalibrates the response to each feature, and maximizes the use of radar data. At the same time, DAF can reduce noise interference in radar data by weighting features, avoiding the loss of secondary details caused by filtering operations, and alleviating the problem of excessive noise in radar data. Finally, due to the influence of the complex weather environment and the model itself, it is difficult for the model to obtain an effective feature representation in the complex weather environment. Therefore, we introduced a batch loss function to enable the model to focus on feature extraction in a complex environment, so as to obtain a more accurate representation of feature information. It reduces model errors and speeds up the convergence of the model. The experiment was conducted on the recently released nuScenes dataset, which provides data records of the entire sensor suite of autonomous vehicles. Experiments prove that our method is superior to other fusion methods.