An Approach to Trajectory Estimation of Moving Objects in the H.264 Compressed Domain

Christian Kaes (1) , Henri Nicolas (1)

1 LaBRI - Laboratoire Bordelais de Recherche en Informatique

This paper presents a simple and fast method for unsupervised trajectory estimation of multiple moving objects within a video scene. It is entirely based on the motion vectors that are present in compressed H.264/AVC or SVC video streams. We extract these motion vectors, perform robust frame-wise global motion estimation and use these estimates to form outlier masks. Motion segmentation on the spatio-temporally filtered outlier masks is performed to detect moving regions in the scene, which are analyzed over time in order to identify similar objects in adjacent frames. The construction of so-called Object History Images (OHIs) is proposed to stabilize the trajectories, which are finally interpolated with X-splines. The system enables real-time analysis with standard hardware.

Domaines: Traitement des images [eess.IV]

Christian Kaes, Henri Nicolas. An Approach to Trajectory Estimation of Moving Objects in the H.264 Compressed Domain. PSIVT, Jan 2009, Tokyo, Japan. ⟨hal-00348684⟩