Abstract:

Road detection is an important part of the environment perception technology of unmanned vehicles. Using computer vision technology to achieve the semantic segmentation of environmental scenes is one of the key technologies to ensure the safe driving of unmanned vehicles. We propose a hybrid auto-encoder semantic segmentation model combining sparse auto-encoder and denoising auto-encoder. Using the sparse semantic encoding of sparse auto-encoder and the robust semantic encoding of denoising auto-encoder makes the features learned by the model more conducive for semantic segmentation. By establishing a reasonable arrangement order and stacking form of the model, an optimal selection of image semantics can be achieved, thereby creating a semantic segmentation model with deep “rich structure”, which can further improve the semantic segmentation performance. Experiments show that this model is simpler with shorter training time and better comprehensive segmentation performance.

Key words: road detection, semantic segmentation, hybrid auto-encoder, rich structure