A hybrid genetic algorithm for optimizing sensing parameters in 3D motion estimation applications

This paper introduces an active vision approach for object motion estimation. The approach is formulated as a problem of controlling the pose of a vision system with the goal of minimizing the uncertainties of the motion estimates. A Kalman filter is employed as the object motion estimation algorithm. The uncertainties of the motion estimates are represented by the variances of the estimates produced by Kalman filter. These variances are updated by a Riccati equation which is constructed as a function of the vision system parameters. A hybrid genetic algorithm is proposed to search for the optimal vision system parameters that minimize the uncertainties of the motion estimates. This hybrid algorithm incorporates the concept of Boltzman's probability from simulated annealing. To improve the speed and accuracy of the proposed algorithm an adaptive gradient based search method is also developed.