Democritus University of Thrace (DUTH)
Ανάπτυξη και υλοποίηση νέων τεχνικών εντοπισμού αυτόνομων ρομποτικών συστημάτων για εφαρμογές πραγματικού χρόνου
Abstract
dc:descriptionVision based navigation systems of autonomous vehicles in unknown terrain, constitute a very active research field that finds applications in other scientific regions as well. In this PhD thesis a number of electronic applications are designed and implemented with the purpose to support the process of simultaneous localization and map building in autonomous vehicles. At the same time computational techniques based on Baysian inference are studied in the framework of a fast and state of the art algorithm for simultaneous localization and mapping, the so called FastSLAM algorithm. These techniques are implemented using an observation system based on stereo vision. Hardware and software implementations are developed with the purpose to localize and build the map in real time. A final result of this thesis is the development of an original electronic circuit with very large scale of integration, for the visual observation and real-time extraction of features, which are appropriate for 3D robotic mapping. The device was developed as a system-on-aprogrammable- chip (SoC), using a Field Programmable Gate Array (FPGA). In this work, feature extraction is based on dense depth maps produced in real-time using a stereo head. For this reason, significant effort was given in order to design and implement hardware systems that are able to accelerate the computational demanding problem of correspondence in stereo pairs. A hardware system-on-a-chip was built which is able to fully parallelize the algorithm Sum of Absolute Differences (SAD), which is a local correlation technique. This hardware system is based on the parallelism of comparisons between a window in the right image and the multitude of candidate windows in the left image. Window-buffering is implemented using on-chip memory. Processing rates of hundreds of frames per second are achieved, and a maximum of 64 levels of disparity is supported. Working further on the problem of accelerating dense depth extraction, we propose a new maximum likelihood technique which can parallelize the steps of dynamic programming. This algorithm can globally optimize correspondences along epipolar scanlines. Based on this technique a new stereo system is proposed and implemented on a programmable chip. This system achieves frame rates similar to those achieved by the local correlation system, however it produces disparity maps of better quality, due to global optimization. Moreover, the system was expanded to receive support from neighboring scanlines, a technique that reduces the error rate. Both systems produce disparities in parallel with the streaming input pixels, so that they output one pixel of disparity for each pair of input pixels of the left and right image frame. Data streaming is controlled by an embedded Nios II processor. The system is supported by external RAM memory and peripheral controllers. A Cyclone II 2C35 medium FPGA device by Altera Corporation was used for the implementation and assessment of the systems. In the final observation system for robot localization and mapping the stereo accelerator based on dynamic programming was used. This stereo detector was supplemented with hardware blocks for feature extraction from the images of the stereo pair. Point features are extracted as corners under the strict condition of consistency between the right and left image. The consistency check is based on the disparity value extracted from the dense depth map. According to the observation system above, the mathematical model of observation is developed, in the framework of the fast localization and mapping algorithm (FastSLAM). This algorithm functions as a filter for uncertainty with respect to observation and vehicle motion. The algorithm is implemented using a general purpose computer, which is a host to the systemon- the-programmable-chip. The host computer and the hardware accelerator are linked through usb2.0 high speed controller. The computational techniques implemented in the framework of FastSLAM algorithm are a Rao-Blackwellized particle filter, which is used to estimate the vehicle track and a multitude of Extended Kalman Filters, used for the position update of the measured landmarks in the map. The total software and hardware system is assessed using a stereo-head developed in the laboratory. The stereo-head is attached to a small radio-controlled platform, appropriate to monitor features in an indoor environment. Part of the system odometry is based on the microcontroller which controls the servomotors, while the orientation of the vehicle is extracted with visual accuracy, making use of successive image frames. The stereo-assisted experiments of simultaneous localization and mapping prove that despite the great computational load, providing support to feature extraction by a hardware accelerator allows the track and map estimation in real-time. The proposed features are lowdimensional but especially robust for 3D mapping, due to the support by the dense depth map. The accurate extraction of vehicle orientation through the proposed visual techniques also contributes to the desired performance
Degree
thesis:*- Grantor dc:publisher
- Democritus University of Thrace (DUTH)
- Year dc:date
- 2010
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Kalomiros, Ioannis
- Καλόμοιρος, Ιωάννης
Subjects
dc:subject × 16- Ηλεκτρονικά συστήματα
- Όραση μηχανών
- Διαμορφούμενο υλικό
- Στερεοσκοπία
- Ρομποτική
- Εντοπισμός και χαρτογράφηση
- Electronic systems
- Machine vision
- Reconfigurable hardware
- Stereo vision
- Robotics
- SLAM ( Simultaneous localization and mapping)
- Επιστήμες Μηχανικού και Τεχνολογία
- Επιστήμη Ηλεκτρολόγου Μηχανικού, Ηλεκτρονικού Μηχανικού, Μηχανικού Η/Υ
- Engineering and Technology
- Electrical Engineering, Electronic Engineering, Information Engineering
Rights
- Language dc:language
- gre
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/25835