Creating Incremental Models of Indoor Environments through Omnidirectional Imaging

• Main Authors: Vicente Román, Luis Payá, Sergio Cebollada, Óscar Reinoso
• Format: Article
• Language: English
• Published: MDPI AG 2020-09-01
• Series: Applied Sciences
• Subjects: mapping; incremental clustering; mobile robots; global-appearance descriptors; omnidirectional images; computer vision
• Online Access: https://www.mdpi.com/2076-3417/10/18/6480

In this work, an incremental clustering approach to obtain compact hierarchical models of an environment is developed and evaluated. This process is performed using an omnidirectional vision sensor as the only source of information. The method is structured in two loop closure levels. First, the Node Level Loop Closure process selects the candidate nodes with which the new image can close the loop. Second, the Image Level Loop Closure process detects the most similar image and the node with which the current image closed the loop. The algorithm is based on an incremental clustering framework and leads to a topological model where the images of each zone tend to be clustered in different nodes. In addition, the method evaluates when two nodes are similar and they can be merged in a unique node or when a group of connected images are different enough to the others and they should constitute a new node. To perform the process, omnidirectional images are described with global appearance techniques in order to obtain robust descriptors. The use of such technique in mapping and localization algorithms is less extended than local features description, so this work also evaluates the efficiency in clustering and mapping techniques. The proposed framework is tested with three different public datasets, captured by an omnidirectional vision system mounted on a robot while it traversed three different buildings. This framework is able to build the model incrementally, while the robot explores an unknown environment. Some relevant parameters of the algorithm adapt their value as the robot captures new visual information to fully exploit the features’ space, and the model is updated and/or modified as a consequence. The experimental section shows the robustness and efficiency of the method, comparing it with a batch spectral clustering algorithm.