Classification of LIDAR Sensor Contaminations with Deep Neural Networks

Classification of LIDAR Sensor Contaminations with Deep Neural Networks
Jyothish K. James, Georg Puhlfürst, Vladislav Golyanik, Didier Stricker
ACM Chapters Computer Science in Cars Symposium (CSCS-2018)

Light detecting and ranging (LIDAR) sensors are extensively studied in autonomous driving research. Monitoring the performance of LIDAR sensors has become significantly important to ensure their reliability and hence guarantee the safety of the vehicle. Underestimation of sensor performance can give away reliable object data, overestimation may result in safety issues. Besides light and weather conditions, the performance is strongly affected by contaminations on the sensor front plate. In this paper, we focus on classifying different types of contaminations using a deep learning approach. We train a deep neural network (DNN) following a multi-view concept. For the generation of training and test data, experiments have been conducted, in which the front plate of a LIDAR sensor has been contaminated artificially with various substances. The recorded data is transformed to contain the essential information in a compact format. The results are compared to classical machine learning techniques to demonstrate the potential of DNN approaches for the problem under consideration.
LIDAR, Deep Learning, Multi-view, Transfer Learning, Scan, Scan Points