Video-based road detection via online structural learning

Video-based road detection is a crucial enabler for the successful development of driver assistant and robot navigation systems. But reliable detection is still on its infancy and deserves further research. In order to adapt to the situation consisting of environmental varieties, an online framework is proposed focusing on exploring the structure cue of the feature vectors. Through the structural support vector machine, the road boundary and non-boundary instances are firstly discriminated. Then they are utilized to fit a complete road boundary. After that, the road region is accordingly inferred and the obtained results are treated as ground truth to update the learned model. Three contributions are claimed in this work: online-learning updating, structural information consideration, and targeted sampling selection. The proposed method is finally evaluated on several challenging videos captured by ourselves. Qualitative and quantitative results show that it outperforms the other competitors.