The recognition of objects and the vehicles surroundings belong to the greatest challenges when it comes to automated driving.

Our solution is to use machine learning methods by mapping complex, not procedurally programmable solutions with the help of neural networks. Therefore, we use a Nvidia Drive PX2, which can process neural networks, is optimized for the application in automotive contexts and has all necessary interfaces.

To enable a vehicle to capture and understand its environment we first train the neural network on the according task. Therefore, we create the network according to the task and optimize it. Then, the data collected in test drives is prepared, annotated, and artificially multiplied to train the network efficiently. Finally, we test, validate and optimize the created network.

This method was applied for example for automated vehicle platoons in order to expand the existing system for the detection of road markings. Standard systems normally use a front camera placed in the center of the vehicle. Platoon driving however has the disadvantage that cameras in such a position see very little, as the vehicle in front takes up most of the space on the picture. Moving the camera to the left enlarges the view, but leaves neural networks puzzled, as the road markings now look skewed due to the different point of view.

Thanks to our testing vehicles we can reproduce such situations, record the relevant data (e.g. camera, radar or lidar) and annotate these, in order to retrain and validate the neural networks.

For annotation purposes we use our own software, which helps us to mark e.g. vehicles, road markings or any type of relevant objects according to your requirements. By adding varying weather conditions we can multiply the data artificially, which enlarges the amount of training material for your neural network.