Fusion of Deep Feature and Hand-Crafted Features for Terrain Recognition

Abstract

Terrain recognition is one of the key problems of mobile robots. It can help the robots understand the surrounding environment. With terrain prediction, the robots could realize autonomous navigation and path planning. This paper focuses on image feature selection for terrain recognition with visions. For terrain recognition tasks, feature is used to represent image information. Traditional visual features can be targeted to express the low-level information like color or texture. The deep feature is extracted by self-learning of neural network, containing richer semantic information than low-level features. There is a complementary relationship between the two. The efficiency and accuracy of terrain recognition is remarkably raised by the fusion of two above features. In the course of algorithm, combination of the off-line training model and on-line recognition model is used to identify the terrain type of the sample, which is to ensure the real-time performance. The corresponding terrain dataset--SDUterrain is established. The algorithm achieves 96% or higher classification accuracy in the experiments based on the SDUterrain Dataset, which is much higher than the single feature classification algorithm.