Mobile robot indoor logical localization method based on scene semantic analysis

A new method of mobile robot indoor logical localization based on scene semantic analysis is proposed. The method characteristic lies in scene modeling using middle-level semantics of scene image to solve the correspondence gap problem between low-level visual features of image and high-level semantics of image, and is applicable to scenes classification and recognition. First, a visual vocabulary is formed by feature clustering using speeded up robust features (SURF). Then pLSA-BoW is utilized to exploit the potential probability distribution of topics in the image modeling. Finally, scene recognition is performed using SVM. There are obvious advantages in computational efficiency using SURF with robust, stability and low-noise, greatly improving the scene recognition speed of the robot. Experiments on mobile reconnaissance robot Hunt-5 designed independently in three types scenes of laboratory, corridor and crossing recognition demonstrate the efficiency of the method with a correct localization rate of 92.5%, satisfing the real-time localization requirement of the mobile robot.