Feature Initialization for Bearing-Only Visual SLAM Using Triangulation and the Unscented Transform

We introduce the direct linear triangulation (DLT) as a general technique for delayed feature initialization for bearing-only visual simultaneous localization and mapping (SLAM). Visual features are tracked over multiple frames, and triangulation is performed to recover a three-dimensional landmark position. The unscented transform is used to estimate the landmark covariance matrix. Our previous work in this field is extended with the novel usage of the DLT, as well incorporating observation information from the current SLAM map into the initialization process, resulting in more accurate landmark estimation. Our technique is simple and efficient, requiring no lengthy Jacobian calculations. We present results demonstrating the technique in a bearing-only SLAM system with a mobile robot.