Traversable region detection based on near-to-far self-supervised incremental learning
Yunlei Chen, Huimin Lu, Junhao Xiao, Hui Zhang · 2014
Effective far-range traversable region detection is a fundamental issue for mobile robots. However, the performance of traditional methods is limited as distance estimation of stereovision system is unreliable beyond 10-15m. In this paper, we proposed a far-range traversable region detection algorithm based on near-to-far self-supervised learning. In the algorithm, superpixel segmentation is employed as preprocessing to reduce the computational complexity. Then, near-range LBP features are extracted for each superpixel. Afterwards, the resulted LBP features are used to train an Incremental Supported Vector Machine (ISVM) for classification, which enhances the far-range region classification performance. Thorough experiments have been carried out utilizing our Nubot mobile robot in outdoor environments. Furthermore the proposed algorithm has also been evaluated using the KITTI Vision Benchmark dataset compared against state of the art algorithms. The results show that the proposed algorithm can detect the traversable regions in a far range efficiently at a high successful rate.