Use this URL to cite or link to this record in EThOS: https://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.754389
Title: Scene understanding by robotic interactive perception
Author: Khan, Aamir
ISNI:       0000 0004 7427 4266
Awarding Body: University of Glasgow
Current Institution: University of Glasgow
Date of Award: 2018
Availability of Full Text:
Access from EThOS:
Access from Institution:
Abstract:
This thesis presents a novel and generic visual architecture for scene understanding by robotic interactive perception. This proposed visual architecture is fully integrated into autonomous systems performing object perception and manipulation tasks. The proposed visual architecture uses interaction with the scene, in order to improve scene understanding substantially over non-interactive models. Specifically, this thesis presents two experimental validations of an autonomous system interacting with the scene: Firstly, an autonomous gaze control model is investigated, where the vision sensor directs its gaze to satisfy a scene exploration task. Secondly, autonomous interactive perception is investigated, where objects in the scene are repositioned by robotic manipulation. The proposed visual architecture for scene understanding involving perception and manipulation tasks has four components: 1) A reliable vision system, 2) Camera-hand eye calibration to integrate the vision system into an autonomous robot’s kinematic frame chain, 3) A visual model performing perception tasks and providing required knowledge for interaction with scene, and finally, 4) A manipulation model which, using knowledge received from the perception model, chooses an appropriate action (from a set of simple actions) to satisfy a manipulation task. This thesis presents contributions for each of the aforementioned components. Firstly, a portable active binocular robot vision architecture that integrates a number of visual behaviours are presented. This active vision architecture has the ability to verge, localise, recognise and simultaneously identify multiple target object instances. The portability and functional accuracy of the proposed vision architecture is demonstrated by carrying out both qualitative and comparative analyses using different robot hardware configurations, feature extraction techniques and scene perspectives. Secondly, a camera and hand-eye calibration methodology for integrating an active binocular robot head within a dual-arm robot are described. For this purpose, the forward kinematic model of the active robot head is derived and the methodology for calibrating and integrating the robot head is described in detail. A rigid calibration methodology has been implemented to provide a closed-form hand-to-eye calibration chain and this has been extended with a mechanism to allow the camera external parameters to be updated dynamically for optimal 3D reconstruction to meet the requirements for robotic tasks such as grasping and manipulating rigid and deformable objects. It is shown from experimental results that the robot head achieves an overall accuracy of fewer than 0.3 millimetres while recovering the 3D structure of a scene. In addition, a comparative study between current RGB-D cameras and our active stereo head within two dual-arm robotic test-beds is reported that demonstrates the accuracy and portability of our proposed methodology. Thirdly, this thesis proposes a visual perception model for the task of category-wise objects sorting, based on Gaussian Process (GP) classification that is capable of recognising objects categories from point cloud data. In this approach, Fast Point Feature Histogram (FPFH) features are extracted from point clouds to describe the local 3D shape of objects and a Bag-of-Words coding method is used to obtain an object-level vocabulary representation. Multi-class Gaussian Process classification is employed to provide a probability estimate of the identity of the object and serves the key role of modelling perception confidence in the interactive perception cycle. The interaction stage is responsible for invoking the appropriate action skills as required to confirm the identity of an observed object with high confidence as a result of executing multiple perception-action cycles. The recognition accuracy of the proposed perception model has been validated based on simulation input data using both Support Vector Machine (SVM) and GP based multi-class classifiers. Results obtained during this investigation demonstrate that by using a GP-based classifier, it is possible to obtain true positive classification rates of up to 80\%. Experimental validation of the above semi-autonomous object sorting system shows that the proposed GP based interactive sorting approach outperforms random sorting by up to 30\% when applied to scenes comprising configurations of household objects. Finally, a fully autonomous visual architecture is presented that has been developed to accommodate manipulation skills for an autonomous system to interact with the scene by object manipulation. This proposed visual architecture is mainly made of two stages: 1) A perception stage, that is a modified version of the aforementioned visual interaction model, 2) An interaction stage, that performs a set of ad-hoc actions relying on the information received from the perception stage. More specifically, the interaction stage simply reasons over the information (class label and associated probabilistic confidence score) received from perception stage to choose one of the following two actions: 1) An object class has been identified with high confidence, so remove from the scene and place it in the designated basket/bin for that particular class. 2) An object class has been identified with less probabilistic confidence, since from observation and inspired from the human behaviour of inspecting doubtful objects, an action is chosen to further investigate that object in order to confirm the object’s identity by capturing more images from different views in isolation. The perception stage then processes these views, hence multiple perception-action/interaction cycles take place. From an application perspective, the task of autonomous category based objects sorting is performed and the experimental design for the task is described in detail.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID: uk.bl.ethos.754389  DOI: Not available
Keywords: QA75 Electronic computers. Computer science
Share: