Visual tracking in occurrence of out-of-field-of-view utilizing background information by particle filter

Major problem of tracking a visual target in occurrence of out-of-field-of-view is unable to obtain information of the target from image frame. We solve this problem by utilizing background information in the image frame. Scenes of tracking a car from an observer in aerial vehicle have been employed. Background information is obtained from road area, building area, and the others. Prior probabilities of existence of the target are pre-assigned for these areas. Two cases have been explored such that 1) road shape as the background information and known observer´s location, and 2) all areas in the image frame as the background information and unknown observer´s location. State space models have been constructed where system model represents time evolution of velocity and posture of the target and observation model represents likelihood computation. State of the target has been estimated sequentially over the image sequence by particle filter. Experiments with synthetic image sequences show the effectiveness of the proposed method.