Comprehensive LiDAR Sensor Noise Model and Application to Texture and Object Classification using Custom Deep Learning

This paper presents a new approach that involves (1) proposing a comprehensive noise model using automotive Light Detection and Ranging (LiDAR) sensor data and (2) applying the model in object classification based on the texture difference of the objects. We first generated our own sensor data from a 16-channel Velodyne LiDAR to measure several objects with different textures. Then, we selected suitable deep learning techniques, with a particular emphasis on data that were not thoroughly considered in conventional analyses, to extract detailed information about intricate textures on the surfaces of different objects. The proposed technique classifies the texture, which in turn facilitates object classification. We tested and validated the proposed method through custom experiments to demonstrate its performance in classifying objects with different textures. The model achieved a high accuracy exceeding 99%, suggesting that our approach has the potential to be applied to classification tasks such as object recognition and segmentation.