Real-Time Three Dimensional Vision

Active systems for collision avoidance in ‘noisy’ environments such as traffic which contain large numbers of moving objects will be subject to considerable interference when the majority of the moving objects are equipped with common avoidance systems. Thus passive systems, which require only input from the environment, are the best candidates for this task. In this paper, we investigate the feasibility of real-time stereo vision for collision avoidance systems. Software simulations have determined that sum-of-absolute-difference correlation techniques match well but hardware accelerators are necessary to generate depth maps at camera frame rates. Regular structures, linear data flow and abundant parallelism make correlation algorithms good candidates for reconfigurable hardware. The SAD cost function requires only adders and comparators for which modern FPGAs provide good support. However accurate depth maps require large disparity ranges and high resolution images and fitting a full correlator on a single FPGA becomes a challenge. We implemented SAD algorithms in VHDL and synthesized them to determine resource requirements and performance. Altering the shape of the correlation window to reduce its height compared to its width reduces storage requirements with negligible effects on matching accuracy. Models which used the internal block memory provided by modern FPGAs to store the ‘inactive’ portions of scan lines were compared with simpler models which used the logic cell flip-flops. From these results, we have developed a simple predictor which enables one to rapidly determine whether a target appliction is feasible.