Optimal parameter selection for intensity-based multi-sensor data registration
Accurate co-registration of multi-sensor data is a primary step in data integration for photogrammetric and remote sensing applications. A proven intensity-based registration approach is Mutual Information (MI). However the effectiveness of MI for automated registration of multi-sensor remote sensing data can be impacted to the point of failure by its non-monotonic convergence surface. Since MI-based methods rely on joint probability density functions (PDF) for the datasets, errors in PDF estimation can directly affect the MI value. Certain PDF parameter values, such as the bin-size of the joint histogram and the smoothing kernel, need to be assigned in advance, since they play a key role in forming the convergence surface. The lack of a general approach to the assignment of these parameter values for various data types reduces both the automation level and the robustness of registration. This paper proposes a new approach for selection of optimal parameter values for PDF estimation in MI-based registration of optical imagery to LiDAR point clouds. The proposed method determines the best parameters for PDF estimation via an analysis of the relationship between similarity measure values of the data and the adopted geometric transformation in order to achieve the optimal registration reliability. The performance of the proposed parameter selection method is experimentally evaluated and the obtained results are compared with those achieved through a feature-based registration method.