MULTI-TARGET CROSS-CAMERA TRACKING METHOD AND RELATED DEVICE

The present invention discloses a multi-target cross-camera tracking method and related device. According to the method, target
detection boxes of multiple tracked targets are identified frame-by-frame from video data in response to tracking target information,
and appearance feature vectors and motion state vectors are extracted; each motion state vector is matched with a target motion submodel for state prediction, to generate state prediction boxes; state prediction boxes are associated with the target detection boxes based on the appearance feature vectors, to generate a local trajectory of each tracked target under corresponding camera; an adjacent camera pair meeting a target transfer condition is selected, and local trajectories are extracted from adjacent camera pair to form a candidate trajectory pair; and a target trajectory pair is selected and is associated to form a global identifier for all tracked targets
based on spatiotemporal transfer probabilities and appearance similarities of all candidate trajectory pairs.