J. Liu, H. Chen, Y. Xu, W. Rong (PRC), T. Vaudrey, and R. Klette (New Zealand)
cell processor, disparity map, computer vision, driver assis tance system, parallel computation, residual image.
Real-time implementations of stereo algorithms are re quired for the latest driver assistance and robotic appli cations. Current speeds of dense stereo algorithms are no where near the required 0.1 seconds per image pair (on a convensional CPU) required for a “real-time” frame rate. This paper describes an efficient parallel implementation of dynamic programming and belief propagation algorithms on a cell processor that speeds up stereo image analysis. Dynamic programming processes image data by scanline optimization; thus it is easily implemented on a cell pro cessor. Belief propagation differs from dynamic program ming by having potentially the whole image area as an area of influence for every pixel; this potentially global opti mization scheme produces improved results, but requires more running time than the dynamic programming method. Furthermore, we define limitations of the Cell architecture for these applications. For evaluation, we use synthetic and real-world image sequences. Real-world images are typi cally degraded by various types of noise, changes in light ing, differing exposures, and so on. Sobel edge and residual images can improve the stereo matching results compared to the use of original real-world images; our results show that a cell processor also reduces running time for these processes.
Important Links:
Go Back
