This study addresses the issue of vision based mobile robot navigation in a partially cluttered indoor environment using a mapless navigation strategy. The work focuses on two key problems, namely vision based obstacle avoidance and vision based reactive navigation strategy. The estimation of optical flow plays a key role in vision based obstacle avoidance problems, however the current view is that this technique is too sensitive to noise and distortion under real conditions. Accordingly, practical applications in real time robotics remain scarce. This dissertation presents a novel methodology for vision based obstacle avoidance, using a hybrid architecture. This integrates an appearance-based obstacle detection method into an optical flow architecture based upon a behavioural control strategy that includes a new arbitration module. This enhances the overall performance of conventional optical flow based navigation systems, enabling a robot to successfully move around without experiencing collisions. Behaviour based approaches have become the dominant methodologies for designing control strategies for robot navigation. Two different behaviour based navigation architectures have been proposed for the second problem, using monocular vision as the primary sensor and equipped with a 2-D range finder. Both utilize an accelerated version of the Scale Invariant Feature Transform (SIFT) algorithm. The first architecture employs a qualitative-based control algorithm to steer the robot towards a goal whilst avoiding obstacles, whereas the second employs an intelligent control framework. This allows the components of soft computing to be integrated into the proposed SIFT-based navigation architecture, conserving the same set of behaviours and system structure of the previously defined architecture. The intelligent framework incorporates a novel distance estimation technique using the scale parameters obtained from the SIFT algorithm. The technique employs scale parameters and a corresponding zooming factor as inputs to train a neural network which results in the determination of physical distance. Furthermore a fuzzy controller is designed and integrated into this framework so as to estimate linear velocity, and a neural network based solution is adopted to estimate the steering direction of the robot. As a result, this intelligent iv approach allows the robot to successfully complete its task in a smooth and robust manner without experiencing collision. MS Robotics Studio software was used to simulate the systems, and a modified Pioneer 3-DX mobile robot was used for real-time implementation. Several realistic scenarios were developed and comprehensive experiments conducted to evaluate the performance of the proposed navigation systems. KEY WORDS: Mobile robot navigation using vision, Mapless navigation, Mobile robot architecture, Distance estimation, Vision for obstacle avoidance, Scale Invariant Feature Transforms, Intelligent framework.

Small, high performance and low cost inspection vehicles, working together as teams of autonomous agents, are well suited to remote inspection tasks in areas that are only accessible through narrow passageways or are hazardous for humans. In particular, the evaluation of complex infrastructures that consist of numerous components with small structural dimensions motivate the application of robotic micro systems for inspection and on-site manipulation. This thesis describes the development of a synergetic multi-robot inspection system for Non-Destructive evaluation (NDE) of engineering structures, enabling magnetic (flux leakage and eddy current), ultrasonic and visual techniques in a complementary fashion. Particular emphasis is placed on the design of the miniature autonomous climbing vehicles, the Eddy Current and Magnetic Flux Leakage payload, the positioning and host system and the fusion of the different NDE data.

No description available.

Road and obstacle analysis are two of the essential building blocks in both Driver Assistance Systems (DAS) and Autonomous Transportation Systems (ATS). Our research focus is to develop computationally efficient algorithm for accurate de- tection of the road boundaries and potential obstacles ba ed on prior knowledge of the highway and urban environments. In this thesis, a novel lane feature extrac- tion algorithm is introduced. It incorporates the global lane shape information to accurately extract feature points that overlap with the lane boundaries. It can be used a a general framework to improve or refine the feature map obtained with a diverse range of local feature extractors. At the lane tracking stage, the performance of the Gaussian Particle Filter (GPF), Gaussian Sum Particle Filter (GSPF) and Sampling Importance Resampling (SIR) particle filter are compared. The GSPF shows a preferable characteristic which is suitable for the lane track- ing application and leads to the best results. For motion-based obstacle detec- tion, we propose a computationally efficient image warping algorithm for motion compensation. This algorithm achieves higher efficiency as well as identical re- sults to perspective mapping based approaches. Furthermore, we investigated stereo vision based obstacle detection and developed a disparity calculation algo- rithm using multi-pass aggregation and local optimisation which utilises the prior knowledge of the traffic scene. This algorithm achieves comparable results to the global optimisation based algorithms with lower computational complexity. Dur- ing the obstacle detection stage, the G-Disparity image. which encloses disparity gradient information, is proposed. Using G-Disparity in conjunction with the -V-Di parity images allows more efficient obstacle extraction with performance improvement over the conventional U- V-Disparity based approaches.

In the domain of a homogenous multi robot system where only single beacon/robot can be activated during the sensor measurement, the problem arises such that which robot should be selected in order to prevent signal interference and more importantly to improve the accuracy of the team position estimation. This research addresses the solution to the problem of a stochastic beacon selection for mobile robot collective localization system. The 'stochastic' and 'collective' terms here differentiate the proposed work from other work. The term 'stochastic' means that beacon selection algorithm does not require a constant or arbitrary topology of sensor/robot formation during the robot navigation. In addition the term 'collective' implies that the shared knowledge among the robot group can be combined and each member can benefit from their team mate resources. The proposed beacon selection algorithm is accomplished by the use of Kalman filter gain as a measure of information between robot relative measurements. The computation of the Kalman gain value and its relation to information quantities contained in the measurements are derived and graphically described through the position and direction of the measurements and the 30" ellipse of the filter a priori error estimates. The preliminary investigations show that the more information results in high Kalman gain and hence high reduction in position uncertainty, i.e. better result in state correction. Therefore the beacon robot whose the set of relative measurements that attains in the highest value of Kalman filter gain, i.e. the most informative measurements, will be selected to activate as a beacon. The proposed stochastic beacon selection algorithm based on Kalman filter gain is then applied to the localization framework based on a recursive estimator, an extended Kalman filter (EKF). The 2D computer based simulations on a fleet of four robots are setup in order to assess the improvement of the fleet localization. The numerical result shows that NRMS errors from EKF are smaller in all states compared to the error from the odometry system by one order. Therefore the EKF localization with relative information significantly improves the accuracy of the odometry measurement. Then the beacon selection algorithm based on Kalman filter gain is added to the simulations. The results show the further improvement of the team localization accuracy over sequential algorithm. The beacon selection strategy based on Kalman filter gain is proved to perform better than sequential algorithm, especially when robots move in random motion. The extension to 3D simulation with the robot group behaviours, cornfield vector and flocking behaviour, also yield the same results obtained from 2D.

This thesis provides techniques to address some outstanding problems in robotic navigation in relatively large environments using simultaneous localisation and mapping (SLAM), resulting in improved competence and reliability of autonomous agents. Autonomous mobile agents are helpful in a diverse range of applications deemed as dull, dirty or dangerous for human operators including mining, defence and underwater explorations. In order to be fully autonomous the robot is required to incrementally construct a map of its vicinity and simultaneously localise itself within the environment based only on its on-board sensory data. However, the robot is confronted with a number of serious challenges that impair large-scale navigation and could even render SLAM results intractable in real time. This work provides efficient strategies for addressing such problems. The techniques presented include an effective method of reducing computational burden of updating the covariance matrix at every step, and cutting down the storage requirements. In addition, the usually complex and tedious task of transforming and fusing sub-maps into a single global map is simplified. Furthermore, an intelligent SLAM (I -SLAM) is introduced that enables accurate place recognition and minimise pose estimation errors. It also provides the means of adaptively adjusting to the nature of natural surface terrains for both indoor and outdoor environments. These techniques promise an enormous potential for autonomous agents operating in unknown environments in terms of consistency, accuracy and efficiency. Simulation results demonstrate the reliability and effectiveness of the system as a means of addressing some of the outstanding challenges with regards to large-scale SLAM performance

Dead reckoning systems are used for navigation of surface robots and these systems commonly rely on conventional odometry such as an encoder counting a number of wheel rotation. This simplicity makes dead reckoning both robust and economic when compared with other navigation systems based, for example, on inertial, landmark recognition or range finding techniques. However, errors can be introduced into dead reckoning systems through wheel slippage or skidding and through manufacturing imperfections of the wheels themselves. These problems of inaccuracy resulting from the use of conventional odometry have been addressed through the design of an optical dead reckoning system for robotic surface navigation. The optical computer mouse has previously been utilised in odometry systems but this has been further developed in an optical dead reckoning system. The use of an optical mouse in a non-contact, dead reckoning system for operation over rough surfaces has been investigated through a number of prototype developments. Systems based on coaxial optical arrangements and telecentric lens technology have been designed and tested. A case study has then been carried out into the application of a telecentric-based optical arrangement for the navigation of an underwater crawling ship hull robot. The prototype units and the case study have demonstrated the suitability of the optical mouse sensor for use in a non-contact odometry system suitable for use over uneven surface profiles.

No description available.

No description available.

No description available.