Unsupervised scene detection for field robots in long-term operation using single camera

Robots´ ability to mine information from the environment, such as traversability and roughness of terrains, is affected not only by the special composition of objects, but also the external factors, such as seasons, weathers and especially lighting conditions. Dividing a long-term operation into certain periods of scenes, where the robot locates in identical environment with unchangeable external conditions, will help the robot mining the knowledge more efficiently for discriminating the growing diversity of the data set. Here we present an unsupervised scene detection system using single camera, based on the assumption that images with similar visual appearances belong to the same scene. The design focuses on how to link an image to a certain scene type, and how to detect scenes starting from an empty dataset. During the operation, a dictionary of the blocks of all images is incrementally constructed, based on which a scene model is also incrementally built to identify the current image as an old scene and to detect a new one. Experimental results show that our system is capable of distinguishing images of different scene types by their visual appearances without any supervision.