Ground Segmentation by CNNs in Velodyne Data
In our recent work, we deployed convolutional neural networks for the task of ground segmentation in sparse Velodyne point cloud data. We have developed a semi-automatic tool for ground annotation (see red points in the figure below) and the transformation of sparse LiDAR data into a dense 2D matrix which can be further processed by CNN. We have designed and trained multiple networks with shallow topologies (3-5 convolutional layers) using the Caffe framework.
Our experiments show, that we have significantly reduced the time requirements of ground segmentation process to the fraction of the time required by the state-of-the-art while preserving the same precision (mAP 99% for annotated frames of KITTI dataset). Processing of a single Velodyne frame by our network using only CPU takes 140 ms on average. Using the GPU acceleration (NVIDIA GeForce GTX 770), the processing time is further reduced to 7ms per frame.