A key component of autonomous outdoor navigation in unstructured environments is the classification of terrain. Recent development in the area of machine learning show promising results in the task of scene segmentation but are limited by the labels used during their supervised training. In this work, we present and evaluate a flexible strategy for terrain classification based on three components: A deep convolutional neural network trained on colour, depth and infrared data which provides feature vectors for image segmentation, a set of exchangeable segmentation engines that operate in this feature space and a novel, air pressure based actuator responsible for distinguishing rigid obstacles from those that only appear as such. Through the use of unsupervised learning we eliminate the need for labeled training data and allow our system to adapt to previously unseen terrain classes. We evaluate the performance of this classification scheme on a mobile robot platform in an environment containing vegetation and trees with a Kinect v2 sensor as low-cost depth camera. Our experiments show that the features generated by our neural network are currently not competitive with state of the art implementations and that our system is not yet ready for real world applications.

abstract: Advancements in computer vision and machine learning have added a new dimension to remote sensing applications with the aid of imagery analysis techniques. Applications such as autonomous navigation and terrain classification which make use of image classification techniques are challenging problems and research is still being carried out to find better solutions. In this thesis, a novel method is proposed which uses image registration techniques to provide better image classification. This method reduces the error rate of classification by performing image registration of the images with the previously obtained images before performing classification. The motivation behind this is the fact that images that are obtained in the same region which need to be classified will not differ significantly in characteristics. Hence, registration will provide an image that matches closer to the previously obtained image, thus providing better classification. To illustrate that the proposed method works, naïve Bayes and iterative closest point (ICP) algorithms are used for the image classification and registration stages respectively. This implementation was tested extensively in simulation using synthetic images and using a real life data set called the Defense Advanced Research Project Agency (DARPA) Learning Applied to Ground Robots (LAGR) dataset. The results show that the ICP algorithm does help in better classification with Naïve Bayes by reducing the error rate by an average of about 10% in the synthetic data and by about 7% on the actual datasets used. / Dissertation/Thesis / M.S. Electrical Engineering 2011

Ground motions from earthquakes are known to be affected by earth's surface topography. Topographic effects are a result of several physical phenomena such as the focusing or defocusing of seismic waves reflected from a topographic feature and the interference between direct and diffracted seismic waves. This typically causes an amplification of ground motion on convex features such as hills and ridges and a de-amplification on concave features such as valleys and canyons. Topographic effects are known to be frequency dependent and the spectral accelerations can sometimes reach high values causing significant damages to the structures located on the feature. Topographically correlated damage pattern have been observed in several earthquakes and topographic amplifications have also been observed in several recorded ground motions. This phenomenon has also been extensively studied through numerical analyses. Even though different studies agree on the nature of topographic effects, quantifying these effects have been challenging. The current literature has no consensus on how to predict topographic effects at a site. With population centers growing around regions of high seismicity and prominent topographic relief, such as California, and Japan, the quantitative estimation of the effects have become very important. In this dissertation, we address this shortcoming by developing empirical models that predict topographic effects at a site. These models are developed through an extensive empirical study of recorded ground motions from two large strong-motion datasets namely the California small to medium magnitude earthquake dataset and the global NGA-West2 datasets, and propose topographic modification factors that quantify expected amplification or deamplification at a site. To develop these models, we required a parameterization of topography. We developed two types of topographic parameters at each recording stations. The first type of parameter is developed using the elevation data around the stations, and comprise of parameters such as smoothed slope, smoothed curvature, and relative elevation. The second type of parameter is developed using a series of simplistic 2D numerical analysis. These numerical analyses compute an estimate of expected 2D topographic amplification of a simple wave at a site in several different directions. These 2D amplifications are used to develop a family of parameters at each site. We study the trends in the ground motion model residuals with respect to these topographic parameters to determine if the parameters can capture topographic effects in the recorded data. We use statistical tests to determine if the trends are significant, and perform mixed effects regression on the residuals to develop functional forms that can be used to predict topographic effect at a site. Finally, we compare the two types of parameters, and their topographic predictive power. / Ph. D.

Today, Sport Utility Vehicles are marketed as both on-road and off-road vehicles. This results in a compromise when designing the suspension of the vehicle. If the suspension characteristics are fixed, the vehicle cannot have good handling capabilities on highways and good ride comfort over rough terrain. The rollover propensity of this type of vehicle compared to normal cars is high because it has a combination of a high centre of gravity and a softer suspension. The 4 State Semi-active Suspension System (4S4) that can switch between two discrete spring characteristics as well as two discrete damper characteristics, has been proven to overcome this compromise. The soft suspension setting (soft spring and low damping) is used for ride comfort, while the hard suspension setting (stiff spring and high damping) is used for handling. The following question arises: when is which setting most appropriate? The two main contributing factors are the terrain profile and the driver’s actions. Ride comfort is primarily dependant on the terrain that the vehicle is travelling over. If the terrain can be identified, certain driving styles can be expected for that specific environment. The terrains range from rough and uncomfortable to smooth with high speed manoeuvring. Terrain classification methods are proposed and tested with measured data from the test vehicle on known terrain types. Good results were obtained from the terrain classification methods. Five terrain types were accurately identified from over an hour’s worth of vehicle testing. Handling manoeuvres happen unexpectedly, often to avoid an accident. To improve the handling and therefore safety of the vehicle, the 4S4 can be switched to the hard suspension setting, which results in a reduced body roll angle. This decision should be made quickly with the occupants’ safety as the priority. Methods were investigated that will determine when to switch the suspension to the handling mode based on the kinematics of the vehicle. The switching strategies proposed in this study have the potential, with a little refinement, to make the ride versus handling decision correctly. Copyright 2007, University of Pretoria. All rights reserved. The copyright in this work vests in the University of Pretoria. No part of this work may be reproduced or transmitted in any form or by any means, without the prior written permission of the University of Pretoria. Please cite as follows: Bester, R 2007, The ride comfort versus handling decision for off-road vehicles, MEng dissertation, University of Pretoria, Pretoria, viewed yymmdd < http://upetd.up.ac.za/thesis/available/etd-10252007-111611 / > / Dissertation (MEng (Mechanical Engineering))--University of Pretoria, 2007. / Mechanical and Aeronautical Engineering / unrestricted

3D semantic segmentation is an expanding topic within the field of computer vision, which has received more attention in recent years due to the development of more powerful GPUs and the newpossibilities offered by deep learning techniques. Simultaneously, the amount of available spatial LiDAR data over Sweden has also increased. This work combines these two advances and investigates if a 3D deep learning model for semantic segmentation can learn to detect terrain roughness in airborne LiDAR data. The annotations for terrain roughness used in this work are taken from SGUs 2D soil type map. Other airborne data sources are also used to filter the annotations and see if additional information can boost the performance of the model.  Since this is the first known attempt at terrain roughness classification from 3D data, an initial test was performed where fields were classified. This ensured that the model could process airborne LiDAR data and work for a terrain classification task. The classification of fields showed very promising results without any fine-tuning. The results for the terrain roughness classification task show that the model could find a pattern in the validation data but had difficulty generalizing it to the test data. The filtering methods tested gave an increased mIoU and indicated that better annotations might be necessary to distinguish terrain roughness from other terrain types. None of the features obtained from the other data sources improved the results and showed no discriminating abilities when examining their individual histograms. In the end, more research is needed to determine whether terrain roughness can be detected from LiDAR data or not.

No description available.

La combinaison de données multi-capteurs joue un rôle croissant dans les systèmes de percep- tion artificielle. Les données de profondeur et les capteurs LiDAR en particulier sont devenus la norme pour les systèmes de vision dans les applications de robotique et de conduite auto- nome. La fusion de capteurs peut améliorer la précision des tâches et a été largement étudiée dans des environnements à ressources élevées, mais elle est moins bien comprise dans les ap- plications où les systèmes peuvent être limités en termes de puissance de calcul et de stockage d’énérgie. Dans l’analyse de la démarche chez l’homme, la compréhension du contexte local de la marche joue un rôle important, et l’analyse en laboratoire à elle même peut limiter la capacité des chercheurs à évaluer correctement la marche réelle des patients. La capacité de classifier automatiquement les terrains de marche dans divers environnements pourrait donc constituer un élément important des systèmes d’analyse de l’activité de marche. Le ter- rain de marche peut être mieux identifié à partir de données visuelles. Plusieurs contraintes (notamment les problèmes de confidentialité liés à l’envoi de données visuelles en temps réel hors appareil) limitent cette tâche de classification au dispositif Edge Computing lui- même, un environnement aux ressources limitées. Ainsi, dans ce travail, nous présentons une architecture de réseau neuronal convolutif parallèle, à fusion tardive et optimisée par calcul de bord pour l’identification des terrains de marche. L’analyse est effectuée sur un nouvel ensemble de données intitulé L-AVATeD: l’ensemble de données Lidar et visibles de terrain de marche, composé d’environ 8000 paires de données de scène visuelles (RVB) et de profondeur (LiDAR). Alors que les modèles formés sur des données visuelles uniquement produisent un modèle de calcul de bord capable d’une précision de 82%, une architecture composée d’instances parallèles de MobileNetV2 utilisant à la fois RVB et LiDAR améliore de manière mesurable la précision de la classification (92%) / Terrain classification is a critical sub-task of many autonomous robotic control processes and important to the study of human gait in ecological contexts. Real-time terrain iden- tification is traditionally performed using computer vision systems with input from visual (camera) data. With the increasing availability of affordable multi-sensor arrays, multi- modal data inputs are becoming ubiquitous in mobile, edge and Internet of Things (IoT) devices. Combinations of multi-sensor data therefore play an increasingly important role in artificial perception systems. Depth data in general and LiDAR sensors in particular are becoming standard for vision systems in applications in robotics and autonomous driving. Sensor fusion using depth data can enhance perception task accuracy and has been widely studied in high resource environments (e.g. autonomous automobiles), but is less well understood in applications where resources may be limited in compute, memory and battery power. An understanding of local walking context also plays an important role in the analysis of gait in humans, and laboratory analysis of on its own can constrain the ability of researchers to properly assess real-world gait in patients. The ability to automatically classify walking terrain in diverse environments is therefore an important part of gait analysis systems for use outside the laboratory. Several important constraints (notably privacy concerns associated with sending real-time image data off-device) restrict this classification task to the edge- computing device, itself a resource-constrained environment. In this study, we therefore present an edge-computation optimized, late-fusion, parallel Convolutional Neural Network (CNN) architecture for the real-time identification of walking terrain. Our analysis is performed on a novel dataset entitled L-AVATeD: the Lidar And Visible wAlking Terrain Dataset, consisting of approximately 8,000 pairs of visual (RGB) and depth (LiDAR) scene data. While simple models trained on visual only data produce an edge-computation model capable of 82% accuracy, an architecture composed of parallel instances of MobileNetV2 using both RGB and LiDAR data, measurably improved classifi- cation accuracy (92%).

This paper deals with the topic of design of a navigation system for an autonomous mobile robot in a park-like environment. Precisely, designing methods for road detection using available sensoric system, designing a mathematical model for fusion of these data, and suggesting a representation of an environment suitable for planning and local navigation.