Stereo vision using a microcanonical mean field annealing neural network

Abstract

A simple stereo algorithm is presented here to obtain the 3D position of a scene's object points. The proposed algorithm can obtain an object point's 3D position by merging the microcanonical mean field annealing network (MCMFA) with the stereo vision system. Comparison with mean field annealing (MFA) or microcanonical simulated annealing (MCSA) reveals that current temperature controls the cooling speed, thereby reducing the computation in MCMFA without degrading the performance. In addition, the initial temperature does not affect the quality of solution in MCMFA because of the new temperature cooling procedure. The correspondence problem is the primary concern of stereo vision. A combinatorial optimization approach is used to resolve the correspondence problem for a set of features extracted from a stereo vision pair. An energy function is defined to represent the solution's constraints and the function is then mapped onto a 2D neural network. Each neuron in the network represents a possible correlation between a feature in the left image and one in the right image. The features are zero-crossing points that are extracted using the LOG (Laplacian of the Gaussian) operator. Zero-crossing points are classified into 16 patterns according to their local connectivity. The difference of the sign value and direction value between a matched pair of zero-crossings can be used to set up the neural node's iteration rule. The network is assumed to be at its stable state when no change occurs in the neurons' state. Finally, the neighbour's disparity threshold (NDT) is used to enhance the precision of the corresponding situation. Once all the corresponding points are found, obtaining the 3D object position is a simple matter of triangulation.