Stereoscopic tracking of bodies in motion

摘要

A fast, highly efficient and robust visual tracking process for multiple moving objects using stereo image sequences taken from a stationary pair of cameras is presented. The algorithm assumes that object motion is restricted to a horizontal plane (e.g. motion of cars on roads or humans walking). Dense stereo image sequences and the visualized locus method1,2 (in which each image sequence is first sampled to produce a 2D spatio-temporal cross-section image) are used to ensure temporal correspondence without search. Edge segments in the left and right spatio-temporal images are then matched. Additional stereo matching constraints are derived by using motion and temporal continuity to reduce the number of ambiguous matches. Speed is achieved by only processing a single spatio-temperal cross-section image from each image sequence. The algorithm succeeds in tracking objects in space and time moving against arbitrarily complex backgrounds and in the presence of occlusion, disappearance and reappearance of object features. The output of the algorithm is the 3D position of moving objects as a function of time.