Affordance-map : learning hidden human context in 3D scenes through virtual human models

Publication Type:
Thesis
Issue Date:
2016
Full metadata record
Files in This Item:
Filename Description Size
01front.pdf119.43 kB
Adobe PDF
02whole.pdf5.05 MB
Adobe PDF
Ability to learn human context in an environment could be one of the most desired fundamental abilities that a robot should possess when sharing workspaces with human co-workers. Arguably, a robot with appropriate human context awareness could lead to a better human robot interaction. This thesis addresses the problem of learning human context in indoor environments by only looking at geometrics features of the environment. The novelty of this concept is, it does not require to observe real humans to learn human context. Instead, it uses virtual human models and their relationships with the environment to map hidden human affordances in 3D scenes. The problem of affordance mapping is formulated as a multi label classification problem with a binary classifier for each affordance type. The initial experiments proved that the SVM classifier is ideally suited for affordance mapping. However, SVM classifier recorded sub-optimum results when trained with imbalanced datasets. This imbalance occurs because in all 3D scenes in the dataset, the number of negative examples outnumbered positive examples by a great margin. As a solution to this, a number of SVM learners that are designed to tolerate class imbalance problem are tested for learning the affordance-map. These algorithms showed some tolerance to moderate class imbalances, but failed to perform well in some affordance types. To mitigate these drawbacks, this thesis proposes the use of Structured SVM (S-SVM) optimized for F1-score. This approach defines the affordance-map building problems as a structured learning problem and outputs the most optimum affordance-map for a given set of features (3D-Images). In addition, S-SVM can be learned efficiently even on a large extremely imbalanced dataset. Further, experimental results of the S-SVM method outperformed previously used classifiers for mapping affordances. Finally, this thesis presents two applications of the affordance-map. In the first application, affordance-map is used by a mobile robot to actively search for computer monitors in an office environment. The orientation and location information of humans models inferred by the affordance-map is used in this application to predict probable locations of computer monitors. The experimental results in a large office environment proved that the affordance-map concept simplifies the search strategy of the robot. In the second application, affordance-map is used for context aware path planning. In this application, human context information of the affordance-map is used by a service robot to plan paths with minimal distractions to office workers.
Please use this identifier to cite or link to this item: