A model for improved association of radar and camera objects in an indoor environment

This paper presents a model for the deviation of distances measured by radar and by optical sensors (3D point clouds). The measured 3D point clouds are typically clustered to objects and the cluster centers are then associated with the radar targets. However, the physical extent and the object geometry cause a divergence of the measured radar range from the cluster centers of the 3D point cloud. Existing solutions either ignore this deviation and compare radar ranges and cluster centers directly without any correction or use object class specific models for the correction. On the contrary, the presented parametric model is independent of the object class. All necessary properties (i.e. object size, flatness and aspect angle) can be obtained online from the covariance matrix of the 3D-point cloud. The model weights are fitted offline using a representative dataset of indoor objects with known radar and cluster correspondences. Furthermore, it is described how this model can be used to localize radar reflections in thermal infrared (TIR) images for a firefighting application. In this particular application, structure-from-motion (SfM) is used to obtain the 3D point cloud from a single moving camera. Experiments with measured data show that the proposed model allows a better data association compared to a direct association.