A multistage filtering technique to detect hazards on the ground plane

We consider the detection of hazards within the ground plane immediately in front of a moving pedestrian. Consecutive views of the scene are acquired by a standard video camera. Using epipolar constraints between views, detected features are matched to compute the camera motion and reconstruct the 3D geometry. Assuming the ground is planar, projective invariance of the cross-ratio and the presence or absence of significant peaks in a Lomb-Scargle periodogram are used in a region-growing technique to label a triangulated mesh as obstructed or unobstructed ground plane. The eventual goal of our approach is to provide an aid to elderly pedestrians in the avoidance of tripping hazards, for example uneven flagstones or kerbs. © 2002 Elsevier Science B.V. All rights reserved.