Pixel material classification using polarimetric LiDAR active imaging.

Active imaging systems, such as light detection and ranging (LiDAR), can return detailed electromagnetic characteristics of each pixel within a field of view. This allows for pixel-by-pixel material identification and classification. Our recently proposed LiDAR system measures the diagonal of the Mueller matrix of each pixel in the field of view, and if there are multiple, temporally distinct reflections in a given pixel, the Mueller matrix from each reflection can be measured. This is accomplished using a time-varying polarization state of the transmitted laser and a two-channel polarization analyzer at the detector. This system has been used in recent work to demonstrate accurate estimation of material Mueller matrices. In this work, we extend the receiver processing to use each estimated Mueller matrix to autonomously perform material identification for scene characterization. The end goal of this manuscript is to explore mathematical techniques of characterizing and classifying pixel surfaces using data from a simulated LiDAR system that incorporates real-world Mueller matrix data. We give an overview of the laboratory-measured dataset and discuss features of the dataset salient to the selection and performance of machine learning algorithms. Classification performance assessment is performed via simulations that incorporate the database of laboratory-measured Mueller matrices. This includes waveform generation, environment simulation, feature extraction, and classification. The simulations show that we can achieve up to 70% classification accuracy on 35 individual classes and 84% accuracy when the data are grouped into five super-classes, provided the assumption of a diagonal Mueller matrix is correct. These results show that the proposed method has promise and should be combined with other methods of classification to achieve even better accuracy.