Semi-supervised Lifelong Learning

Automatic road segmentation from live video streams is a crucial prerequisite for various robotics applications such as path following or automatic return from teleoperated mission in case of signal loss. Such system often needs to operate under a wide spectrum of challenging operational conditions regarding climate (diverse light conditions, like in direct sunlight, overcast, sunset), and surrounding environment (segment both, high-quality roads as well as roads barely visible even for humans including sand, concrete, tarmac, gravel). Pre-trained models usully work only in conditions similar to a training set and domain adaption is still an unsolved problem. Instead, we treat the video input as an infinite stream of training data and perform on-the-fly life-long learning of our model.
We tackle the problem of a robust detection of shady and highlighted roads from a monocular camera by a self-supervised learning algorithm combining the frequency based vanishing point estimation and probabilistic texture segmentation. Hence, our method does not require any additional sensor such as laser range finder, stereo camera, etc. to determine the training area. A combination of the two is advantageous since these methods are complementary, e.g. in situations like sudden road texture or illumination change, the probabilistic model for texture segmentation is not consistent with current road surface but it is possible to use the detected vanishing point until the probabilistic model is adapted.

Publications

Robust Detection of Shady and Highlighted Roads for Monocular Camera Based Navigation of UGV
Miksik O., Petyovsky P., Zalud L. and Jura P.
In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA) 2011, Shanghai, China
PDF | Show BibTex | Show Details | Poster | Errata


Rapid Vanishing Point Estimation for General Road Detection
Miksik O.
In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA) 2012, St. Paul, USA
PDF | Show BibTex | Show Details


Adapting Polynomial Mahalanobis Distance for Self-supervised Learning in an Outdoor Environment
Richter M., Petyovsky P. and Miksik O.
In Proceedings of the IEEE International Conference on Machine Learning and Applications (ICMLA) 2011, Honolulu, USA
PDF | Show BibTex | Show Details | Poster