Good Features for SLAM

Abstract: SLAM is an overdetermined online estimation problem whose size grows with time, as new measurements are collected with each new image. Most highly overdetermined problems can be converted to smaller overdetermined problems without suffering a loss in accuracy. In that spirit, good features for SLAM investigates the question: How can we determine what visual features to keep or to remove to preserve accuracy?

If we can answer the question above, then we can start to bound the computational time of SLAM for real-time operation in large-scale and uncertain environments. Our first stab at this question was the 2015 CVPR paper “Good Features to Track for Visual SLAM” ¹ Since this was our first foray into SLAM after having studied 3D reconstruction and point cloud processing, it was applied to an EKF-based SLAM method (1pt RANSAC EKF-SLAM). Filter-based, indirect SLAM algorithms are the easiest and quickest to get running. However, for robustness they typically require RANSAC or some other outlier rejection method. Here, we employed the observability conditioning of the SLAM estimation problem as a filter for the RANSAC process. Rather than randomly selecting points, the algorithm rank orders the points according to a conditioning score, then tests the resulting hypotheses. The scoring here is the observbility condition number associated to the minimum singular value of the observability matrix for the discrete time SLAM process. For more efficient calculations, a special version of the observability matrix is computed (the stripped observability matrix) whose rank and spectral properties agree with the observability matrix. The better the observability conditioning a point provides, the higher ranking it should have. Selecting points based on this criteria provides a mechanism to be robust to error because observability conditioning is related to estimation convergence and error rejection.

The approach was intended to provide proof-of-concept for the idea that the features being tracked could be actively managed and incorporated into pose estimation based on how they influence or condition the SLAM estimation. As a concept, there are many ways to implement the approach. Furthermore, it was meant to complement existing SLAM solutions. Thus we anticipated that good features as a concept connected to solution conditioning would naturally translate to other SLAM solution.