Vizualna odometrija u stvarnome vremenu za robusnu navigaciju u nepoznatim složenim okruženjima

doctoral thesis

• Title (english): Real-time visual odometry for robust navigation in unknown complex environments
• Author: Igor Cvišić
• Mentor: Ivan Petrović (mentor)
• Committee member: Ivan Petrović (predsjednik povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; (Department of Control and Computer Engineering); Zagreb
• Defense date and country: 2018, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Electrical Engineering; Automation and Robotics
• Universal decimal classification (UDC): 621.3 - Electrical engineering

Abstract

U ovome doktorskom radu opisani su algoritmi lokalizacije mobilnog agenta stereo kamerom. Razvoj algoritama bio je motiviran natjecanjem EuRoC, Izazov 3: Nadzor i održavanje postrojenja mikro-letjelicom. Prvo je opisan novorazvijeni aloritam vizualne odometrije SOFT (engl. Stereo Odometry with Feature Tracking), koji drži prvo mjesto na rang listi podatkovnog skupa KITTI već dulje od dvije godine. Algoritam SOFT zasnovan je na pomnom odabiru značajki i njihovom praćenju kroz vrijeme podpikselskom preciznošću. Također, SOFT kombinira monokularnu metodu za određivanje rotacije kamere sa stereo metodom za određivanje translacije kamere u svrhu povećanja robusnosti na nesavršeno umjeren stereo par. Nadalje, u radu je opisano proširenje vizualne odometrije na SOFT-SLAM (engl. Stereo Odometry with Feature Tracking - Simultaneous Localization and Mapping). Algoritam SOFT-SLAM oslanja se na visoku točnost vizualne odometrije, dok globalnu konzistenciju karte osigurava optimizacijom grafa položaja kamere. Razvijen je s ciljem izvođenja u stvarnome vremenu na računalu s ograničenim resursima. Stoga je organiziran u dvije glavne dretve - sinkronu dretvu vizualne odometrije koja osigurava rad u stvarnome vremenu i asinkronu dretvu kartiranja i zatvaranja petlje za globalnu konzistenciju. Algoritam SOFT-SLAM izvodi se s konstantnim periodom izvođenja 20Hz na dvije jezgre računala. Na natjecanju EuRoC, Izazov 3, Etapa IIa - Testiranje, Zadatak 2: Estimacija stanja i točnost upravljačkog sustava, SOFT-SLAM je postigao najviše bodova za točnost u konkurenciji pet finalista natjecanja. Također, u radu su uspoređeni rezultati algoritma SOFT-SLAM s rezultatima dvaju vodećih algoritama u području - ORB-SLAM2 i LSD-SLAM, i to na dvama relevantnim podatkovnim skupovima - EuRoC i KITTI. Usporedba pokazuje da u prosjeku najtočnije rezultate postiže SOFT-SLAM, uz korištenje znatno manje računalnih resursa. Nadalje, u radu je opisan razvoj nove inercijske stabilizacijske platforme za percepcijske senzore na višerotorskoj letjelici. Za razliku od standardnih rješenja, razvijena inercijska stabilizacijska platforma uspješno stabilizira senzore velikih raspona mase, pri čemu učinak stabilizacije raste proporcionalno masi senzora. Razvijena platforma značajno smanjuje utjecaj dinamike letjelice na točnost lokalizacije letjelice i kartiranja prostora.

Abstract (english)

In this thesis we describe algorithms for localization of mobile agents using a stereo camera. The development of algorithms was motivated by the EuRoC competition, Challenge 3: Plant Servicing and Inspection with micro aerial robots. First, we describe the newly developed visual odometry algorithm dubbed stereo odometry with feature tracking (SOFT), which has been holding the first place on the rankings table of the KITTI dataset for more than two years. The SOFT algorithm is based on a careful feature selection and tracking through time in subpixel resolution. Moreover, SOFT combines rotation estimation using a monocular camera with stereo based translation estimation in order to improve robustness to imperfect stereo calibration. Furthermore, in the thesis we also present the extension of the SOFT odometry to simultaneous localization and mapping (SOFT-SLAM). The SOFT-SLAM algorithm relies on the high accuracy of the visual odometry, while the global map consistency is ensured by optimizing the camera pose graph. The approach was developed with the goal of being executed on an embedded computer with limited resources. Given that, it is organized in two separate threads: the synchronous thread for visual odometry operating in real-time and asynchronous thread for mapping and loop closing to ensure global consistency. SOFT-SLAM runs in constant time at 20 Hz frequency on two processor cores. In the EuRoC competition, Challenge 3, Stage IIa - Benchmarking, Task 2: State Estimation and Control System Accuracy, among the five competing finalist SOFT-SLAM achieved the highest score in localization accuracy. In the thesis, we also compare the proposed algorithm to two state-of-the-art approaches, namely ORB-SLAM2 and LSD-SLAM, on two relevant datasets, namely the EuRoC and KITTI dataset. Comparison showed that on average SOFT-SLAM achieves the highest score using significantly less computing resources. Furthermore, in the thesis we also describe development of a new inertial stabilization platform for perception sensors to be mounted on a multirotor flying robot. As opposed to standard solutions, the developed inertial stabilization platform successfully stabilizes sensors largely varying in weight, where the effect of stabilization is proportional to the weight of the sensor. The developed platform significantly reduces the effects of flight dynamics on the accuracy of localization and mapping of the flying robot.

• Keywords: vizualna odometrija; stereo kamera; lokalizacija; SLAM; inercijska stabilizacija
• Keywords (english): visual odometry; stereo camera; localization; SLAM; inertial stabilization
• Language: croatian
• URN:NBN: urn:nbn:hr:168:234676
• Study programme: Title: Electrical Engineering and Computing; Study programme type: university; Study level: postgraduate; Academic / professional title: Doktor znanosti elektrotehnike i računarstva (Doktor znanosti elektrotehnike i računarstva)
• Type of resource: Text
• Extent: 126 str. ; 30 cm.
• File origin: Born digital
• Access conditions: Closed access
• Created on: 2019-04-03 15:32:23