A realtime stereo vision system design using distinctive similarity measure

The stereo algorithm based on distinctive similarity measure shows better matching capability due to matching based on the distinctiveness of a point, however it may be difficult to achieve a realtime performance due to excessive computations. In this paper, we propose a realtime stereo vision system with a comparable matching rate. We separate rows and columns when computing the adaptive support-weights, so that we can reuse the intermediate data. Also, we increase the performance by utilizing 18 processing elements in parallel and efficiently schedule the input-output timing for the intermediate data. We minimize the critical path delay by applying the pipelining to division and the simple approximation to exponential functions. The proposed architecture has been designed with VerilogHDL and synthesized with DB Hitek 110nm standard cell library. It shows 565MHz (56fps) of the maximum frequency with 74K gates and 2.4KB of memory while showing 2.11% of the error rate for the tsukuba image. © The Korean Institute of Electrical Engineers