Sparse Camera Positioning

Accurately estimating the pose of large arrays of fixed indoor cameras is a long-standing challenge in computer vision, since most existing methods depend on cameras sharing overlapping fields of view. This work presents a methodology that positions fixed cameras with and without overlapping views in complex indoor scenarios.

A subset of fiducial markers, printed on regular paper, is strategically placed and progressively relocated through the space. An additional mobile camera records these markers as it moves, gradually connecting every fixed camera in the network, even when consecutive cameras never see each other directly. A final optimization stage minimizes the reprojection error of all observed markers while enforcing real-world physical constraints, such as camera and marker coplanarity.

The method was validated on novel datasets built specifically for this work, spanning corridors, floors, and an entire multi-story building, and it outperforms prior sparse-positioning techniques, establishing a new state of the art for camera networks that cannot rely on overlapping views.

A small set of fiducial markers, printed on regular paper, is placed within view of one fixed camera and then physically relocated, one hop at a time, into view of the next. A moving camera follows this trail and continuously records the markers as they travel, capturing the local geometric relationship between each pair of fixed cameras along the way, even when those two cameras never actually see one another.

Chaining every one of these pairwise relationships across the whole space builds a single connected graph linking all fixed cameras. A final bundle-adjustment-style optimization stage then refines that graph jointly, minimizing the reprojection error of every observed marker while enforcing real-world physical constraints such as camera and marker coplanarity, to recover every camera's pose in one consistent coordinate frame.