A DSP multiprocessor architecture for live camera calibration

This paper presents a system for real-time calibration of moving and zooming video cameras. In order to satisfy the real-time constraints, the calibration of each video frame has to be computed before the acquisition of the next frame is completed (40 milliseconds for PAL standard). Because of such time requirements, the proposed system is based on a multi-DSP processor architecture: the digitized video signal is processed by a DSP system hosted in a standard PC. The computed calibration parameters are transferred to the application running on the host, which uses the camera parameters for real time digital video effects, such as live mixing of the acquired video stream with synthetic sequences. Compared to existing systems, the proposed technique represents a low-cost solution, as it is based on a simple hardware architecture. Moreover, differently from most existing systems, the proposed technique does not need additional hardware for camera calibration, such as mechanical encoders or auxiliary cameras, but just needs some markers arbitrarily placed in the scene. The proposed tracking and calibration technique has been implemented and optimized for running in parallel on a multiple-DSP architecture. Experimental tests show that the overall computing time has never been greater than 13 milliseconds, thus satisfying the required real-time constraints.