Category-specific higher layer of a hierarchical spatial model

A compact and expressive spatial representation is crucial for effective operation of a service robot in the real world environments. We present here:

  • A new category-specific higher layer of a hierarchical representation of space that is based on 2D laser-range data.
  • A new algorithm (MCABE) that finds a dictionary of exemplar elements from a multi-category set, based on the affinity measure defined among pairs of elements.

The freely available Domestic Rooms Dataset has been extended in this work and used for model evaluation in the context of a room categorization problem.

Local-Map Visibility Graphs:

Higher layer parts, named local-map visibility graphs (LMVGs) are formed by merging partial views of the environment, represented by lower layers. Merging is performed using a simultaneous localization and mapping (SLAM) algorithm that produces local maps, from which graph-based structures are extracted (Figure 1).

Figure 1: LMVG creation pipeline. (a) A schematic view of an example room with robots short-path marked with red. The scheme is shown only for visualisation, data used in our work was obtained in real-world environments. (b) Schematic view of three example scans obtained along the short-path. (c) Three partial views of the room represented by lower-layer parts, acquired at different positions along the short-path. These representations, as well as all the following images, were derived from real-world data. (d) All partial views merged into a local map. (e) Visibility-graph with all the visible connections.

MCABE:

Part selection in the higher layer construction is performed by the new learning algorithm (Figure 2), named Multi-Category Affinity-Based Exemplars-search (MCABE). The method is general and is not restricted to our model.


Figure 2: The MCABE algorithm finds a dictionary of exemplar elements from a multi-category set, based on the affinity measure defined among pairs of elements.

Some Results:

Room categorization is performed through higher-layer part detection. Some results, obtained on the Domestic Rooms Dataset, are shown below in Table 1. The proposed approach achieved state-of-the-art results by achieving the most balanced categorization performance over all categories.

Table 1: A confusion matrix showing the results of the proposed method on the Domestic Rooms Dataset. Category abbreviations: LR-living room, CO-corridor, BA-bathroom, KI-kitchen, BE-bedroom, WC-toilet.

Publications for the topic of higher layer of a hierarchical spatial model