Recognition of Point Sets Objects in Indoor Scenes

With the wide application of 3D intelligent sensing technologies such as Lidar and Stereo Camera in robotics and driverless driving field, the acquisition of point cloud data is getting easier, and the cost is becoming lower. The research on point cloud data is gradually transiting from lower-level geometric feature extraction to higher-level semantic understanding. Due to the disorder of point cloud data format, most researchers will transform point cloud data to regular 3D voxel grids, collections of images, depth maps, etc. which will inevitably lead to huge data processing problems. In this paper, based on the indoor scene, we design a new neural network to process point cloud data, which solves the problem of disorder and rotation invariance of point cloud data. Theoretically, this network structure shows strong performance. In experiment, there is an accuracy rate of 89.7% on the test set, this method is superior to current mainstream methods. Experiments show that the improved network structure can accurately identify objects in indoor scenes and has strong robustness.