From vision to actions: Towards adaptive and autonomous humanoid robots

Although robotics research has seen advances over the last decades robots are still not in widespread use outside industrial applications. Yet a range of proposed scenarios have robots working together, helping and coexisting with humans in daily life. In all these a clear need to deal with a more unstructured, changing environment arises. I herein present a system that aims to overcome the limitations of highly complex robotic systems, in terms of autonomy and adaptation. The main focus of research is to investigate the use of visual feedback for improving reaching and grasping capabilities of complex robots. To facilitate this a combined integration of computer vision and machine learning techniques is employed. From a robot vision point of view the combination of domain knowledge from both imaging processing and machine learning techniques, can expand the capabilities of robots. I present a novel framework called Cartesian Genetic Programming for Image Processing (CGP-IP). CGP-IP can be trained to detect objects in the incoming camera streams and successfully demonstrated on many different problem domains. The approach requires only a few training images (it was tested with 5 to 10 images per experiment) is fast, scalable and robust yet requires very small training sets. Additionally, it can generate human readable programs that can be further customized and tuned. While CGP-IP is a supervised-learning technique, I show an integration on the iCub, that allows for the autonomous learning of object detection and identification. Finally this dissertation includes two proof-of-concepts that integrate the motion and action sides. First, reactive reaching and grasping is shown. It allows the robot to avoid obstacles detected in the visual stream, while reaching for the intended target object. Furthermore the integration enables us to use the robot in non-static environments, i.e. the reaching is adapted on-the- fly from the visual feedback received, e.g. when an obstacle is moved into the trajectory. The second integration highlights the capabilities of these frameworks, by improving the visual detection by performing object manipulation actions.

Enhancing human action recognition with region proposals

Deep convolutional network models have dominated recent work in human action recognition as well as image classification. However, these methods are often unduly influenced by the image background, learning and exploiting the presence of cues in typical computer vision datasets. For unbiased robotics applications, the degree of variation and novelty in action backgrounds is far greater than in computer vision datasets. To address this challenge, we propose an “action region proposal” method that, informed by optical flow, extracts image regions likely to contain actions for input into the network both during training and testing. In a range of experiments, we demonstrate that manually segmenting the background is not enough; but through active action region proposals during training and testing, state-of-the-art or better performance can be achieved on individual spatial and temporal video components. Finally, we show by focusing attention through action region proposals, we can further improve upon the existing state-of-the-art in spatio-temporally fused action recognition performance.

R1SVM: A randomised nonlinear approach to large-scale anomaly detection

The problem of unsupervised anomaly detection arises in a wide variety of practical applications. While one-class support vector machines have demonstrated their effectiveness as an anomaly detection technique, their ability to model large datasets is limited due to their memory and time complexity for training. To address this issue for supervised learning of kernel machines, there has been growing interest in random projection methods as an alternative to the computationally expensive problems of kernel matrix construction and sup-port vector optimisation. In this paper we leverage the theory of nonlinear random projections and propose the Randomised One-class SVM (R1SVM), which is an efficient and scalable anomaly detection technique that can be trained on large-scale datasets. Our empirical analysis on several real-life and synthetic datasets shows that our randomised 1SVM algorithm achieves comparable or better accuracy to deep auto encoder and traditional kernelised approaches for anomaly detection, while being approximately 100 times faster in training and testing.

Elliptical summary randomisation for sensor-based human activity recognition

Many conventional statistical machine learning al- gorithms generalise poorly if distribution bias ex- ists in the datasets. For example, distribution bias arises in the context of domain generalisation, where knowledge acquired from multiple source domains need to be used in a previously unseen target domains. We propose Elliptical Summary Randomisation (ESRand), an efficient domain generalisation approach that comprises of a randomised kernel and elliptical data summarisation. ESRand learns a domain interdependent projection to a la- tent subspace that minimises the existing biases to the data while maintaining the functional relationship between domains. In the latent subspace, ellipsoidal summaries replace the samples to enhance the generalisation by further removing bias and noise in the data. Moreover, the summarisation enables large-scale data processing by significantly reducing the size of the data. Through comprehensive analysis, we show that our subspace-based approach outperforms state-of-the-art results on several activity recognition benchmark datasets, while keeping the computational complexity significantly low.

Extending visual perception with haptic exploration for improved scene understanding

Scene understanding has been investigated from a mainly visual information point of view. Recently depth has been provided an extra wealth of information, allowing more geometric knowledge to fuse into scene understanding. Yet to form a holistic view, especially in robotic applications, one can create even more data by interacting with the world. In fact humans, when growing up, seem to heavily investigate the world around them by haptic exploration. We show an application of haptic exploration on a humanoid robot in cooperation with a learning method for object segmentation. The actions performed consecutively improve the segmentation of objects in the scene.

Texture for Script identification

The problem of determining the script and language of a document image has a number of important applications in the field of document analysis, such as indexing and sorting of large collections of such images, or as a precursor to optical character recognition (OCR). In this paper, we investigate the use of texture as a tool for determining the script of a document image, based on the observation that text has a distinct visual texture. An experimental evaluation of a number of commonly used texture features is conducted on a newly created script database, providing a qualitative measure of which features are most appropriate for this task. Strategies for improving classification results in situations with limited training data and multiple font types are also proposed.