Systems composed of multiple sensors for exteroceptive perception are becoming increasingly common, such as mobile robots or highly monitored spaces. However, to combine and fuse those sensors to create a larger and more robust representation of the perceived scene, the sensors need to be properly registered among them, that is, all relative geometric transformations must be known. This calibration procedure is challenging as, traditionally, human intervention is required in variate extents. This paper proposes a nearly automatic method where the best set of geometric transformations among any number of sensors is obtained by processing and combining the individual pairwise transformations obtained from an experimental method. Besides eliminating some experimental outliers with a standard criterion, the method exploits the possibility of obtaining better geometric transformations between all pairs of sensors by combining them within some restrictions to obtain a more precise transformation, and thus a better calibration. Although other data sources are possible, in this approach, 3D point clouds are obtained by each sensor, which correspond to the successive centers of a moving ball its field of view. The method can be applied to any sensors able to detect the ball and the 3D position of its center, namely, LIDARs, mono cameras (visual or infrared), stereo cameras, and TOF cameras. Results demonstrate that calibration is improved when compared to methods in previous works that do not address the outliers problem and, depending on the context, as explained in the results section, the multi-pairwise technique can be used in two different methodologies to reduce uncertainty in the calibration process.

Multimodal approaches in perception bring redundancy and enhance the robustness of solutions for object detection and recognition. This is applicable in diverse situations ranging from the industrial process up to intelligent vehicles. However, to fuse or merge information from multiple sensors it is necessary to know their relative placement, also known as spatial registration, or extrinsic calibration. This work proposes a solution to register infrared and visual cameras seamlessly in an automatic way based on the simultaneous perception of a specific target, namely a large-sized ball. Images are used in grayscale for all cameras, and a Fast R-CNN deep learning-based architecture is used to detect fast and accurately the position of the target in a set of image frames. Sets of 3D ball centers generate point clouds whose matching provides the geometric transformation that registers each pair of cameras. As the detection results have a high degree of confidence (around 98%) the automatic registration of the cameras is also achieved with very good precision as demonstrated in the results.