Simultaneous pose and correspondence determination using Differential Evolution

This paper proposes an Differential Evolution based solution for the problem of pose estimation of 3D objects using 2D images, without any correspondence information. The procedure consists of looking for proper rotation and translation parameters, such that the projections of a set of object points best fit points extracted from the 2D image. First, the problem is modeled as a pose estimation problem whose solution gives the best matching between the object points and the image points, and a simplified object function is chosen. Then, it is solved with the Differential Evolution algorithm, enhanced with a new efficient scheme. Randomized crossover probability is also used to improve the performance as well as a new technique to handle boundary constraints for periodic parameters such as rotation euler angles. Experimental results with synthetic camera images validate the advance of the proposed algorithm compared to the existing Genetic Algorithm based solution, and show that its computational efficiency is also much higher.