• Home
  
  
• Übersicht
• Forschung
• Lehre
• Mitarbeitende
• Publikationen
• Auszeichnungen
• Sonstiges
  
  
• Aktuelles
• Kontakt

GET-Forschungsseminar Abstracts

Toward a Robust Method for Monocular SLAM

Hamed Piarehzadeh Zeidani, GET Lab

23.09.2015, 16:30, P 1.4.17

Abstract:

The problem of autonomous navigation in an unknown environment is currently an active and challenging field of robotics. SLAM (Simultaneous Localization And Mapping) is the key mean toward solving this problem. SLAM algorithms are basically working with sensory data to build the surrounding map and estimate the moving body pose at each moment. Among all type of available sensors, camera based SLAM methods become more and more popular thanks to the recent developments of better and yet more affordable cameras. Pure visual SLAM (PVSLAM) which refers to the SLAM methods using only camera images as sensory data. When PVSLAM uses only one single camera, it is called monocular SLAM in contrast to stereo SLAM in which two cameras are available. The challenging problem of monocular SLAM algorithms is the landmark depth retrieval in depth motion due to the non-Gaussianity of depth. A known solution to this problem is inverse depth parametrization method, which works well in case of side motion; whereas, it diverges if the camera experiences motion in depth. In a work done recently in GET Lab, we have used relative camera motion estimation based on a modified 8-point method (proposed in GET Lab) along with the inverse depth method. The results showed significant improvement in motion estimation since the modified 8-point method provided a good initial estimation for camera motion parameters, while EKF-SLAM corrected the scale to some extent. Despite the fact that results were promising in a context of long term navigation, the problem of landmark depth estimation was still a key factor to limit the accuracy of algorithm in scale drift correction. In a further attempt to address this issue in SLAM, it will be tried to use another measurement model based on epipolar geometry inside EKF-SLAM. This master degree thesis will be an effort toward learning an EKF monocular SLAM method with better performance in motion estimation. The main focus of this work will be on landmark tracking between frames, recovering the depth in depth motion, scale drift correction and the speed of the complete process.