Speaker Tracking in Seminars by Human Body Detection

Abstract. This paper presents evaluation results of a method for tracking speakers in seminars from multiple cameras. First, 2D human tracking and detection is done for each view. Then, 2D locations are converted to 3D based on the calibration parameters. Finally, cues from multiple cameras are integrated in a incremental way to refine the trajectories. We have developed two multi-view integration methods, which are evaluated and compared on the CHIL speaker tracking test set. 1 Task and Data Set The task in this evaluation exercise is to track the 3D head locations of a speaker in seminars. In practice, only the ground plane projections of the 3D head locations are used to evaluate the performance. The test set contains 24 segments captured from the UKA site and 2 segments from the ITC site. For each segment, four side view cameras and one optional top-down camera are used to record the seminar. Camera calibration information including radial distortion, is provided. Each video contai...

• Type: Conference
• Year: 2006
• Where: CLEAR
• Authors: Bo Wu, Vivek Kumar Singh, Ram Nevatia, Chi-Wei Chu
• Keywords: 3D Head Locations, Biometrics, Camera Calibration Information, CLEAR 2006, Multiple Cameras