The camera placement project focuses on the automatic selection of optimal camera configurations (camera locations, orientations etc.) for multi-camera networks. Previous approaches largely focus on proposing various objective functions to achieve different tasks. Most of them suffer from 2 major drawbacks: First, most of them are objective-specific meaning that the approach cannot be easily applied to other objectives. Second, most of them do not generalize well to large-scale networks. For integer programing based approaches, the scale of the search space is limited by the optimization methods. For heuristic based one, no optimality of any sort can be ensured. To tackle these drawbacks, we propose a statistical framework of the problem as well as propose a Trans-Dimensional Simulated Annealing (TDSA) algorithm to effectively deal with it.
The current system is able to deal with the following objectives:
- Full coverage of 2D floorplan while minimising the number of cameras.
- Redundant coverage of critical regions while minimising the number of cameras.
- Achieve user specified face size while minimising the number of cameras needed. The specified face size ensures that the captured face images are large enough for various automatic or manual applications such as criminal identification.
- Combinations of the above objectives.
Current progress involves translating the 2D environment into full 3D so that partial occlusions, such as furniture, can be correctly dealt with.
PhD Student