The training of autonomous driving algorithms requires huge amount of data. The data labeled by human cost too much labor and the quality is not guaranteed. Overfitting can be alleviated by combining deep neural network groups, so large data set can be annotated through a few labeled data. And the annotation accuracy can be improved by human’s post-processing.
The application of multi-sensor for inference and computing to enhance the intelligence and robustness of the unmanned vehicles is popular. The data is collected from Lidar, traditional cameras, millimeter wave radar, infrared cameras, multi-spectral cameras, etc.
The rare events in the training data which are difficult to collect has great impact on the reliability of the algorithms. The platform combined the real data and the virtual models to generate realistic rare events. So that the trained autonomous driving algorithms has a better performance in handling emergent conditions.