Optimization-based automatic parameter tuning for stereo vision

Stereo vision is an important 3D sensing technique for producing dense point clouds required for robotic navigation and manipulation. It can provide excellent depth resolution at high frame rates and is potentially smaller, cheaper and consumes less power than systems using active sensor devices, due to its use of standard imaging components like cameras. However, stereo vision system can generate high quality point clouds only when its parameters are appropriately tuned. To tune these parameters manually is not only tedious but also challenging, due to the large number of parameters and their non-linear effect on the depth map quality. In this paper, we present an optimization-based method to automatically tune the stereo parameters. In particular, we first adjust the disparity range to ensure the entire scene can be covered in the resultant depth map, and then use non-linear optimization to refine other parameters for the optimal depth map quality. Our tuning process is efficient and can update adaptively according to changing environment. Experiments on the teleoperation tasks using the Atlas robot validate our approach, and demonstrate the improvement it brings for the teleoperation effectiveness.