A Semi-Automated Multi-Sensor Data Labeling Process for Deep Learning in Automotive Radar Environment

The continuous improvement of computing power in recent years is supporting Artificial Intelligence (AI) to achieve a breakthrough, not merely in the development of automated vehicles. AI algorithms are used to extract patterns or features from a large amount of sensor data that are used to support the driving task. The number of environment detection sensors required in automated vehicle will increase significantly from SAE Level 3. Especially with the limited range of electric vehicles, an energy-intensive analysis of the recorded data is considered to be disadvantageous, which should be optimized by using appropriate neural networks. Our work aims at providing a blueprint for the process of creating labeled datasets for the development of neural networks for pattern recognition in radar data in the automotive environment. We present various methods for ensuring data quality, i.a. ensemble techniques for the automated labeling, and manual post-processing utilizing quantitative quality measures. The results are to be evaluated under consideration of defined test scenarios, which are primarily oriented towards inner-city intersection scenes. Using an exemplary dataset, the process of data acquisition and semi-automated labeling will be discussed in terms of their performance.