ScenE Similarity Recognition with Integrated High-Dimensional Point Cloud Features

To enhance the scene understanding capability of inspection robots in dynamic environments, this paper proposes a system for calculating scene similarity based on point cloud data to comprehend the connections between two scenes. The point cloud data is acquired through ORB-SLAM2 and integrated with YOLOv5 to exclude dynamic interference, ensuring the precise capture of scene data. The similarity calculation system downsamples the input point cloud, uses the Iterative Closest Point (ICP) algorithm for preprocessing and alignment of the point clouds, and adapts to variable observation conditions through point cloud enhancement techniques. Feature extraction employs self-attention mechanisms and Convolutional Neural Networks (CNNs) for dimensionality enhancement to capture the details of the points, and finally calculates the similarity of the feature vectors using random evaluation metrics. Experimental results show that in the TUM dataset and actual scene tests, the similarity of data collected multiple times in the same scene exceeds 90%, while the similarity in different scenes is below 70%.