Face tracking with an efficient use of the ME engine in a video processor

Face tracking is a widely used technique in image processing and computer vision. To implement face tracking without significant increase of hardware cost, recent research proposes a hardware-reuse algorithm that utilizes the DCT (discrete cosine transform) and ME (motion estimation) engines included in video processors. When the hardware-reuse algorithm is applied to a real MPEG-4 codec processor called Vidan, the success rate of face tracking is severely degraded due to mismatch of the search window size. The paper proposes a new technique to map the hardware-reuse algorithm onto the Vidan processor. To reduce computation time while achieving high success rate, the appropriate image size is chosen to be 64×64. Since the search window size of the Vidan ME engine is 48×48, the image is decomposed by four subimages with the size of 48×48. With each subimage given one by one as the input, the proposed algorithm runs the ME engine four times. To support the new mapping technique, a FIFO queue with 281 entries is added to the existing ME engine. Simulation results show that the proposed algorithm achieves a near optimal success rate.