An Integration Model of Blind Spot Estimation and Traversable Area Detection for Indoor Robots

Traversable area detection is essential for autonomous robot navigation, enabling robots to identify safe areas, plan optimal paths, and avoid obstacles. In complex indoor environments, this task is particularly challenging due to various obstacles and dynamic changes. In this study, an improved traversable area detection model for indoor robots is proposed, which integrates blind spot estimation through multilevel image refinement and normal enhancement. The proposed model segments ground areas traversable for robots while identifying blind spots that are currently invisible but may become visible. The proposed method acquires red-green–blue (RGB) and depth images from an RGB depth (RGB-D) camera, fills depth data holes, and uses a deep learning network for floor, wall, and ceiling segmentation. It then estimates the room structure, renders the main structure’s depth values, and compares them with the filled depth values to identify initial foreground obstacles. Next, a normal maps enhanced image gradient (NMEIG) module is presented to enhance the RGB images, providing a foundation for further refinement. Then, a semantic global edge refinement (SGER) module is designed to adjust the global segmentation results by combining the initial traversable areas and foreground obstacles. In addition, a sliding window local refinement (SWLR) module is presented to further distinguish between foreground and background on a finer scale. Finally, the background parts within the room’s ground structure are identified as the traversable areas, and the foreground obstacle parts are estimated as the blind spot areas. Experimental results demonstrate that the proposed method achieves high accuracy and robustness in indoor traversable area detection with blind spot estimation, providing a reliable and intelligent solution for indoor mobile robots in autonomous scene understanding.