ABSTRACT
Descriptors play an important role in point cloud registration. The current state-of-the-art resorts to the high regression capability of deep learning. However, recent deep learning-based descriptors require different levels of annotation and selection of patches, which make the model hard to migrate to new scenarios. In this work, we learn local registration descriptors for point clouds in a self-supervised manner. In each iteration of the training, the input of the network is merely one unlabeled point cloud. Thus, the whole training requires no manual annotation and manual selection of patches. In addition, we propose to involve keypoint sampling into the pipeline, which further improves the performance of our model. Our experiments demonstrate the capability of our self-supervised local descriptor to achieve even better performance than the supervised model, while being easier to train and requiring no data labeling.
ABSTRACT
Using camera sensors for ground robot Simultaneous Localization and Mapping (SLAM) has many benefits over laser-based approaches, such as the low cost and higher robustness. RGBD sensors promise the best of both worlds: dense data from cameras with depth information. This paper proposes to fuse RGBD and IMU data for a visual SLAM system, called VINS-RGBD, that is built upon the open source VINS-Mono software. The paper analyses the VINS approach and highlights the observability problems. Then, we extend the VINS-Mono system to make use of the depth data during the initialization process as well as during the VIO (Visual Inertial Odometry) phase. Furthermore, we integrate a mapping system based on subsampled depth data and octree filtering to achieve real-time mapping, including loop closing. We provide the software as well as datasets for evaluation. Our extensive experiments are performed with hand-held, wheeled and tracked robots in different environments. We show that ORB-SLAM2 fails for our application and see that our VINS-RGBD approach is superior to VINS-Mono.
