Trinocular visual odometry for divergent views with minimal overlap

We present a visual odometry algorithm for trinocular systems with divergent views and minimal overlap. Whereas the bundle adjustment is the preferred method for multi-view visual odometry problems, it is infeasible if the number of features in the images-such as in HD videos-is large. We propose a divide and conquer approach, which reduces the trinocular visual odometry problem to five monocular visual odometry problems, one for each individual camera sequence and two more using features matched temporally from consecutive images from the center to the left and right cameras, respectively. Unlike the bundle adjustment method, whose computational complexity is O(n³), the proposed approach allows to match features only between neighboring cameras and can therefore be executed in O(n²). Assuming constant motion of the cameras, temporal tracking therefore allows us to make up for the missing overlap between cameras as objects from the center view eventually appear in the left or right camera. The scale factors that cannot be determined by monocular visual odometry are computed by constructing a system of equations based on known relative camera pose and the five monocular VO estimates. The system is solved using a weighted least squares scheme and remains over-defined even when the camera path follows a straight line. We evaluate the resulting system using synthetic and real video sequences that were recorded for a virtual exercise environment.