Risk analysis for smart homes and domestic robots using robust shape and physics descriptors, and complex boosting techniques

In this paper, the notion of risk analysis within 3D scenes using vision based techniques is introduced. In particular the problem of risk estimation of indoor environments at the scene and object level is considered, with applications in domestic robots and smart homes. To this end, the proposed Risk Estimation Framework is described, which provides a quantified risk score for a given scene. This methodology is extended with the introduction of a novel robust kernel for 3D shape descriptors such as 3D HOG and SIFT3D, which aims to reduce the effects of outliers in the proposed risk recognition methodology. The Physics Behaviour Feature (PBF) is presented, which uses an object's angular velocity obtained using Newtonian physics simulation as a descriptor. Furthermore, an extension of boosting techniques for learning is suggested in the form of the novel Complex and Hyper-Complex Adaboost, which greatly increase the computation efficiency of the original technique. In order to evaluate the proposed robust descriptors an enriched version of the 3D Risk Scenes (3DRS) dataset with extra objects, scenes and meta-data was utilised. A comparative study was conducted demonstrating that the suggested approach outperforms current state-of-the-art descriptors.