Indoor point cloud recognition with deep convolutional networks

摘要

With the development of laser radar technology, more and more fields have begun to use laser radar to acquire 3D point cloud information. The crux and premise of 3D object recognition and 3D model semantic segmentation is the depth feature of 3d point cloud. Therefore, it is significant for indoor intelligent robots to recognize 3D objects by using laser radar. However, unlike the regular arrangement of pixels in 2D images, the 3D point cloud data is irregular and disordered, which means it is difficult to acquire local related information between the 3D point cloud with direct convolution operation. At present, the research focus of 3D object recognition is the method based on deep learning. At this stage, the deep convolutional neural network constructed by PointConv can achieve a high level in the semantic segmentation of 3D point cloud. First, this paper introduces a model named PointConv. To balance the performance and complexity of the model, this paper simplifies the PointConv which called Mini-PointConv to reduce the occupation of network computing resources while ensuring the accuracy of the model segmentation results. Furthermore, the method of ScanNet is adopted to test the Mini-PointConv, which shows that the improved network has achieved a good experimental result in 3D scene semantic segmentation tasks and gained a better performance as balance as well. Finally, the Mini-PointConv is tested in a variety of indoor environments using laser radar and obtain a good indoor 3D point cloud recognition result.