Towards real time 3D tracking and reconstruction on a GPU using Monte Carlo simulations

This paper addresses the problem of camera tracking and 3D reconstruction from image sequences, i.e., the monocular SLAM problem. Traditionally, this problem is solved using non-linear minimization techniques that are very accurate but hardly used in real time. This work presents a highly parallelizable random sampling approach based on Monte Carlo simulations that fits very well on the graphics hardware. The proposed algorithm achieves the same precision as non linear optimization, getting real time performance running on commodity graphics hardware. Both accuracy and performance are evaluated using synthetic data and real video sequences captured with a hand-held camera. Moreover, results are compared with an implementation of Bundle Adjustment showing that the presented method gets similar results in much less time.