In automated driving, we require precise environment models. These models need to include not only information about occupied or free space, but also semantic information such as drivable lanes, sidewalks or different classes of traffic participants. At our institute, we already developed detectors for vehicles, cyclists and pedestrians using Deep Learning on occupancy grid maps from range sensor data. However, to further improve classification accuracy and extend the set of classes (e.g. the detection of different lanes), we aim to also include camera data into the inference process.
In this work the environment model should be represented by a multi-layer grid map, including range-sensor occupancies, free-space, reflectance as well as RGB camera images mapped to the ground surface. This way, we can not only transfer labeled training data from camera images to the scale-independent grid map domain but also include additional features from range sensors in our learning process.
At first you should search and discuss related work on semantic classification of camera images and grid maps. Based on your findings, you may implement a recent method or a new approach for semantic classification. This approach should be evaluated using the KITTI dataset. Finally, your method might be validated on our experimental vehicle.
We are happy to answer questions regarding the topic, reference literature oder alternative topics.