Visual-inertial SLAM has been a contemporary research theme with various emerging commercial applications like robot navigation, augmented reality, 3D mapping etc. With the advent of several SLAM systems the theory of multiview geometry has been put to practical use. A typical SLAM system consists of several sub-systems including: visual-odometry, sensor-fusion, place recognition backend, place recognition frontend, posegraph solver. For a successful commercial deployment of SLAM algorithm it is important that the SLAM system be failsafe. In this thesis, we present several techniques for fail-safety of a SLAM system. We start by proposing an edge based visual-odometry method. The advantage of edge based visual odometry over traditional methods based on corner features and optic...[ Read more ]

Visual-inertial SLAM has been a contemporary research theme with various emerging commercial applications like robot navigation, augmented reality, 3D mapping etc. With the advent of several SLAM systems the theory of multiview geometry has been put to practical use. A typical SLAM system consists of several sub-systems including: visual-odometry, sensor-fusion, place recognition backend, place recognition frontend, posegraph solver. For a successful commercial deployment of SLAM algorithm it is important that the SLAM system be failsafe. In this thesis, we present several techniques for fail-safety of a SLAM system. We start by proposing an edge based visual-odometry method. The advantage of edge based visual odometry over traditional methods based on corner features and optical flow is that such methods also work well in featureless human built environment like corridors. Another advantage is that, the proposed method has a large convergence basin which allows for more reliable odometry computation under large motion or low frame rates. Next we present a learning based whole-image descriptor for loop detection. We demonstrated much higher recall rates compared to existing bag-of-visual-words based loop detection methods. Unlike previous loop detection methods which only evaluate their methods on fronto-parallel scenes, we tested our on datasets involving large viewpoint difference. In addition to higher recall, our method involves an order of magnitude less model storage size compared to bag-of-words dictionary and also an order of magnitude lesser FLOPS (floating point operations) making it suitable for a realtime SLAM system. We also propose a robust feature matching scheme and a local bundle optimization based computation for reliably estimating relative pose at loop detections. Unlike some existing works which merge trajectories from multiple runs offline, we develop a pose graph solver which is able to keep track of multiple co-ordinate systems, identify and recover from kidnaps live and in realtime. Extensive online experimental results are presented throughout the thesis. We conclude by proposing future research opportunities.