Ground Mobile Robot Traversable Area Recognition Based on 3D CNN and Attention Mechanism

Abstract

As robot technology advances, ground mobile robots have been widely used in dangerous operation environments such as search, rescue, and bomb disposal. The generation of accurate and reliable traversable area maps is crucial for the autonomous navigation of ground mobile robots, especially in rough, uneven, and complex environments. To acquire a more accurate robot traversable area, this paper presents a traversable area identification method based on 3D convolution and attention mechanism. This method uses the 3D LiDAR point cloud information obtained by laser radar scanning. Firstly, it voxelizes the point cloud to improve the identification efficiency. Then, it optimizes the sparse Bird's Eye View Network to identify the voxelized point cloud. A residual module is introduced between the sub-manifold convolution layer and sparse convolution layer of the 3D sparse convolution to solve the training difficulty caused by network depth. In addition, an attention mechanism module is added to the end of point cloud feature extraction to allow the network to pay more attention to useful information in the features. Experimental results show that our network can achieve superior identification results compared to the Bird's Eye View Network.