We use KITTI Dataset which contains Lidar data, generated by a data collection vehicle from KIT, Karlsruhe, Germany. First, we tried to preprocess the data on our own by cleaning intensities, reduce dimensions and classes, which went pretty well. Afterwards, we wanted to feed our own CNN/ANN wich processed data. Unfortunately, we were not able to build a working Network on our own by several reasons. Afterwards, we decited to shift strategy and use one of the well-known state-of-the-art networks on the KITTI dataset. Unfortunately, Vote3Deep network did not run on our engines because of an unknown error. Due to not finding mandatory packaged despite being installed, VoxelNet did not work as well. Later, we could successfully train PointsNet which is based on VoxelNet and achieved an accuracy of about 58%. If we had some more time to train, we could have achieved much better results.
Furthermore, we generated value by an business idea to establish trust of customers to automotive technologies. However, the app will also be usefull in terms of labelling cars, pedestrians and other classes in images. We tried to develop an education platform, where customers can learn about "How the technology is working" but also has the chance to play against the algorithm during a journey in an autonomous driving car.
Link to used PointsNet: https://github.com/charlesq34/frustum-pointnets