Efficient Feature Parameterisation for Visual SLAM Using Inverse Depth Bundles

Flexibility and robustness of visual SLAM systems have been shown to benefit from an inverse depth parameterisation of features. However the increased number of 6 parameters per feature presents a problem to real-time EKF SLAM implementations because their computational complexity scales quadratically with the size of the state vector. Recent work tackles this for instance by converting the representation of well-established features from inverse to regular depth. In this paper, we propose a parameterisation where bundles of features share a common representation of the view-point they were initially observed from. According to the experiments performed, a feature occupies effectively about 1.5 state parameters in the proposed approach, allowing real-time performance for maps with more than 200 features.