Efficient Ground Truth Labelling of Lidar Data via Deep Learning

Autoren

Pavel Jiroutek, Jan Olšina, Miroslav Zima,
Valeo Autoklimatizace k.s., Praha - Strašnice

Zusammenfassung

The process of ground truth labelling of reference data is inevitable, yet demanding and expensive task in the development of autonomous driving systems. The presented approach uses state of the art machine learning techniques for object detection, tracking and classification applied on the annotation process for statistical validation. Experimental results of the automated annotation process performance and efficiency on lidar data are presented. One of the main benefits of the outlined approach is its easy adaptability to different sets of annotation rules, different datasets of reference data and its ability to gradually increase the output quality when more data is available. It is an essential step towards robust development of fully autonomous driving (AD) systems and significant cost reduction of the validation process.

Mitglieder-Login