Recognizing substantially occluded objects in confined spaces is a very challenging problem for ground-based persistent
surveillance systems. In this paper, we discuss the ontology inference of occluded object recognition in the context of
in-vehicle group activities (IVGA) and describe an approach that we refer to as utilization-based object recognition
method. We examine the performance of three types of classifiers tailored for the recognition of objects with partial
visibility, namely, (1) Hausdorff Distance classifier, (2) Hamming Network classifier, and (3) Recurrent Neural Network
classifier. In order to train these classifiers, we have generated multiple imagery datasets containing a mixture of
common objects appearing inside a vehicle with full or partial visibility and occultation. To generate dynamic
interactions between multiple people, we model the IVGA scenarios using a virtual simulation environment, in which a
number of simulated actors perform a variety of IVGA tasks independently or jointly. This virtual simulation engine
produces the much needed imagery datasets for the verification and validation of the efficiency and effectiveness of the
selected object recognizers. Finally, we improve the performance of these object recognizers by incorporating human
gestural information that differentiates various object utilization or handling methods through the analyses of dynamic
human-object interactions (HOI), human-human interactions (HHI), and human-vehicle interactions (HVI) in the context
of IVGA.
|