DROID-NeXt-SLAM: ConvNeXt-Enhanced Method for Indoor Visual Localization and Mapping

Visual SLAM aims to simultaneously estimate the motion of a camera and reconstruct the 3D structure of an unknown environment in real-time using visual data. Although end-to-end learning SLAM or VO system methods show impressive potential, most existing models suffer from problems such as insufficient utilization of geometric constraints and overfitting, which seriously limit the performance of SLAM or VO systems. Based on the DROID-SLAM framework, this paper tries to effectively utilize geometric constraints of the SLAM scenario and fully exploit the hybrid global and local features during feature extraction and matching, thus propose a pyramid-based method with improved ConvNext structure. Depthwise separable convolutions are used to replace traditional convolutions, significantly reducing the number of model parameters and computational complexity. Large kernel convolutions are used to expand the receptive field, simulating the global attention mechanism of Transformer to capture global information and reduce optimization bias. Furthermore, a pyramid-based feature fusion design is introduced to enhance the utilization of hierarchical correlation information, improving robustness to scale variations and occlusions. Experiments show that the proposed method achieves better localization accuracy. The average ATE on the entire fr1 series of the TUM-RGBD dataset demonstrates a 28 % improvement compared with DROID-SLAM.