Robust Visual Techniques For Robotics Based On Machine Learning

Recent advances in robotics has advanced the field in great magnitudes in recent years with various application requirements being satisfied in the industry,at home,in hospitals,in the military,but to mention a few.Vision for most robotics,quite similarly to traditional computer vision,is faced with numerous challenges.In the realm of modern day robotics however,vision systems have become a cross-disciplinary field which combines machine vision and machine learning into a single framework.Perhaps,the advent of machine learning has been one of the greatest advances in technology in recent years which has offered new ways by which robots can be designed with smart,robust and highly adaptive vision and real-time interaction capabilities.In studying the problem of robotic vision therefore,machine learning is a key component that cannot be foregone.By observing the domains within which robots are deployed,it becomes clear that speed,robustness and adaptability are one of the highest requirements that can be placed on the design of robotic vision and learning systems.The ability to detect,track,recognize,classify and understand real life objects are some of the core requirements of modern day robotics.These behavioural primitives establish the foundation upon w hich more complex systems such as Human Computer Interaction(HCI),Robot Assisted Living(RAL),and other high level operations can be realized.With great advances being achieved in imaging systems and learning techniques,modern-day robotics still struggles to fulfil the basic fore-mentioned problems due to visual challenges including occlusion,illumination variations,clutter,pose variations,deformation,rotations(in-plane and out-of-plane)and drift.Relatively earlier approaches have proposed to tackle these problems by abstracting visual learning as 2-dimensional in nature(2D)but recent advances in imaging and availability of higher computational platforms have allowed the problems of robotic vision and learning to be addressed in 3-dimensional(3D)frameworks.The work in this thesis studies vision and learning for robotics in a deep and extensive nature.We offer manifold contributions in multiple areas associated with the problem.1.Firstly,we address the object detection and tracking problems in real-time.As opposed to previous schemes,we propose 3D robust real-time tracking schemes which do not require offline modelling but are able to learn from the target's behaviour online.Our tracking schemes adopt model-less learning frameworks that are capable of learning the dynamics of a target in real-time using a robust feature-learning framework which combines RGB and Depth features to form an RGBD feature space.2.Secondly,our proposed robust tracking schemes effectively harness circulance and dense sampling in realizing highly robust and impressively fast real-time tracking.A feedback strategy is incorporated into the tracking scheme for robust learning.This makes our trackers highly feasible for real-time systems such as robotics.The kernel circulance also allows our trackers to be easily scaled for single target and multi-target applications.Our RGBD tracking schemes overcome the occlusion,deformation and illumination variation problems.3.Furthermore,we propose a target prediction algorithm for tracking based on Distributed Particle Filtering(DPF).This algorithm which we term DCIDPF is able to achieve high-speed multiple target tracking with minimal particles and is capable of modelling the interaction between multiple targets.This predictive tracking technique overcomes the drift,object occlusion and target prediction problems.4.Towards real-time object classification,we propose a novel multi-class classification scheme based on Ada Boost-RVM for industrial robotics applications.The scheme is proposed with novel feature extraction schemes based on Beamlets and Wavelet transforms and achieves high multi-class classification accuracies.The algorithm overcomes overfitting and degradation due to its robust learning with Random Vector Machines(RVMs)and Adaptive Boosting(Ada Boost).5.Finally,we proposed a real-time 3D object detection and recognition scheme for pose-invariant object recognition.The scheme works with RGBD features which are robust to illumination variations.For this algorithm,we introduce an Improved Triplet Loss Function(TLF)for Multi-Channel Convolutional Neural Network(MC-CNN)training in a fast and accurate manner.Our proposed Triplet CNN-based Learning Scheme is capable of recognizing multiple object poses in real-time and overcomes the clutter and object pose variation problems.All our proposed visual learning schemes for robotics are verified two-fold;Firstly,benchmarking experiments with state-of-the-art are performed and secondly,real-time verification experiments are performed using the 6-legged wall-climbing robot.The validation approaches demonstrate the effectiveness,robustness and superiority of our schemes.