Reconfigurable Computing Architecture for Accurate Disparity Map Calculation in Real-Time Stereo Vision

This paper presents a novel hardware architecture using FPGA-based reconfigurable computing (RC) for accurate calculation of dense disparity maps in real-time, stereo-vision systems. Recent stereo-vision hardware solutions have proposed local-area approaches. Although parallelism can be easily exploited using local methods by replicating the window-based image elaborations, accuracy is limited because the disparity result is optimized by locally searching for the minimum value of a cost function. Global methods improve the quality of the stereo-vision disparity maps at the expense of increasing computational complexity, thus making real-time application not viable for conventional computing. This problem becomes even more evident when stereo vision is a single step integrated into a more complete image elaboration flow, where the depth maps are used for further detection, recognition, stereo reconstruction, or 3D enhancement processing. Our approach exploits a parallel and fully pipelined architecture to implement a global method for the calculation of dense disparity maps based on the dynamic programming optimization of the Hamming distance of the Census-transform cost function. The resulting stereovision core produces results that are significantly more accurate than existing hardware solutions using FPGAs that are based upon local approaches. The design was implemented and evaluated on an Altera Stratix-III E260 FPGA in a GiDEL PROCStar-III board. Tests were performed on 640×480 stereo images, with a Census transform window size = 3, correlation window size = 5, and disparity ranges of 30 and 50. Our hardware architecture achieved a speedup of about 319 and 512 respectively for the two disparity ranges, when compared to an optimized C++ implementation executed on a 2.26 GHz Xeon E5520 core. High accuracy in the output disparity map, together with high performance in terms of frames per second, make the proposed architecture an ideal solution for 3D robot-assisted medical systems, tracking, and autonomous navigation systems, where accuracy and speed constraints are very stringent.