In this project, an observability function (e.g. volume of the field of view FOV, error of 3D reconstruction) is optimized by repositioning and reorienting multiple cameras with respect to a predefined mounting area (e.g. the ceiling of the room).
In order to solve the optimal camera placement problem, an optimization method needs to be chosen that finds the optimal parameters iteratively. However, the properties of the objective function decrease the number of suitable optimization methods: The volume of the FOV alone is a non-continuous, non-convex function, and more complex vision algorithms usually adopt the behavior. Furthermore, in case the objective function is estimated by Simulation, the objective function is computationally expensive and it is a black-box, which means that operations other than function evaluations (e.g. gradient) are hardly possible or even more expensive.
The video below illustrates the iterative improvement of vision system when using three different types of optimization methods. The objective value is compared to the non-iterative, heuristical placement of the cameras in the corners of the ceiling (green).
Special thanks goes to the advise of Prof. J. Pannek and to J. Völkel’s contributions.