신경회로망 기반 스테레오 비젼을 이용한 실시간 거리 계측

A real time depth measurement scheme with a neural network-based stereo vision is proposed. The depth information is very useful for autonomous vehicle or intelligent robot tasks. Image matching and distance interpolation are necessary for the stereo vision based depth computation. Computation load is a problem for real time applications. The proposed Lateral Information Propagation Neural Networks (LIPN) is for the real time stereo vision. Hardware systems for dynamic programming and interpolation can be implemented with the LIPN for image matching and for depth information of whole image, respectively. The networks are composed of two layers: input and output layers. Each output unit has connections from outputs of its neighbor units as well as that of its input unit. The output value of each unit is updated toward the weighted sum of inputs and outputs of some neighbor units at each time step. The information of each unit propagates to its neighbor units with such neurocomputing mechanism. The proposed idea is proven through some simulation for the stereo images.