An Efficient 3D Point Cloud-Based Place Recognition Approach for Underground Tunnels Using Convolution and Self-Attention Mechanism

Existing place recognition methods overly rely on effective geometric features in the data. When directly applied to underground tunnels with repetitive spatial structures and blurry texture features, these methods may result in potential misjudgments, thereby reducing positioning accuracy. Additionally, the substantial computational demands of current methods make it challenging to support real-time feedback of positioning information. To address the challenges mentioned above, we first introduced the Feature Reconstruction Convolution Module, aimed at reconstructing prevalent similar feature patterns in underground tunnels and aggregating discriminative feature descriptors, thereby enhancing environmental discrimination. Subsequently, the Sinusoidal Self-Attention Module was implemented to actively filter local descriptors, allocate weights to different descriptors, and determine the most valuable feature descriptors in the network. Finally, the network was further enhanced with the integration of the Rotation-Equivariant Downsampling Module, designed to expand the receptive field, merge features, and reduce computational complexity. According to experimental results, our algorithm achieves a maximum score of 0.996 on the SubT-Tunnel data set and 0.995 on the KITTI data set. Moreover, the method only consists of 0.78 million parameters, and the computation time for a single point cloud frame is 17.3 ms. These scores surpass the performance of many advanced algorithms, emphasizing the effectiveness of our approach.