Precision lateral control for highway automation

Purpose and goal

The PRELAT project has addressed robustness of lateral positioning for lateral control of automated driving. The target application is automation of trucks for driving on highways. The project is focused on lateral control for which information on drivable road surface and lane markings are of primary interest. In the early autopilot demonstrators proved to have unsatisfactory reliability for a continuous operation such as lane following. The two main ideas pursued in the project have been adding lidar to the camera and deploying convolutional neural networks for the fusion.

Expected results and effects

The PRELAT project has successfully contributed to an area of rapid expansion in the research community and in the automotive industry. The robustness of perception for lateral control has improved. The Volvo Group have explored and refined the results, and will continue to do so for the expected introduction of more automated driving vehicles. A licentiate thesis has been written and the PhD student at Chalmers is proceeding to a doctoral thesis that will be completed soon.

Planned approach and implementation

The research question was: how can we add other sensor modalities and how can we quantify the improvement. The choice was to evaluate lidar point clouds and to combine it with camera based vision. The main method was to study perception as a component in order to evaluate its performance. The technical approach has been supervised learning, where a human annotates features in an image, e.g. lane markings, signs, cars, etc. Evaluation has been performed off-line. On the wish list was data from our own vehicles, but it proved to costly to take for the project alone.