Data Protection Regulation Compliant Dataset Generation for LiDAR-Based People Detection Using Neural Networks

The use of LiDAR sensor technology for people detection offers a significant advantage in terms of data pro-tection. In LiDAR point clouds, unlike camera images, people can be detected but not identified without further information. LiDAR sensors are, therefore, particularly suitable for detecting people in publicly accessible places and reacting accordingly to the number of people, for example, with on demand services at airports. Due to the anonymity of people in LiDAR point clouds, personal data is protected, and approval for implementing such a detection system is simpler than that of comparable camera systems. In this paper, we present a measurement setup that covers the configuration of the sensor setup, the creation of a dataset for training neural networks for object detection, and the object detection itself. The measurement setup generates an average of 2408 automatically labeled point clouds per sensor, per hour. The SECOND network trained with this dataset achieves average precision for the intersection over union of the 2D view with a threshold of 0.5 of 87.67 %, the PV-RCNN of 85.74 % and an average precision for the average orientation similarity with a threshold of 0.5 of 89.56 %, and for the PV-RCNN of 87.81 %.