A visual dirt detection system for mobile service robots

Basic service robot technologies become more and more mature so that applications far beyond fetch and carry tasks will be possible in near future. Using a multifunctional service robot as a cleaning assistant is one interesting and challenging application. To be of any use, cleaning robots have to be able to verify the results of their work or have to find dirt in the beginning for point-wise cleaning. This paper presents an approach towards the detection of various dirt on textured surfaces using a low-cost RGB-D sensor. The algorithm aims at removing the regular part of the image with spectral residual filtering leaving only the dirt and some noise. With further post-processing it is possible to localize many kinds of dirt in the image. The proposed method is fast, reliable, quite robust to illumination changes and does not require any prior training of the surfaces to inspect.