Incremental scene detection in outdoor environment based on hierarchical bag-of-words model

For the purpose of autonomously carrying out tasks in diversified environments, robots are required to possess the ability of scene understanding, and scene detection is considered as one of its most important components. Due to the continuity of time and space in a specific scene, it is hypothesized that mobile robots remain in the same scene during one period of time, and the image sequences from the same scene share a similar visual appearance. Therefore, an incremental scene detection method that requires no prior knowledge is proposed. By establishing the connection between images and scenes through a hierarchical bag of words (BoW) model, our method makes scene detection more similar to human cognitive process. Firstly, every image that is captured by robots in real time is segmented into blocks. Secondly, a dynamic clustering algorithm is implemented to incrementally build the low-level dictionary, according to which features of the high-level BoW model are extracted. Then another dynamic clustering algorithm is implemented for incremental scene detection, so that the current image is classified as either an experienced scene, or an unexperienced scene, until a new scene is detected. Experimental results show that our method can effectively complete autonomous scene detection without prior knowledge. © 2020, Editorial Department of Control Theory & Applications South China University of Technology. All right reserved.