A Binary Descriptor Invariant to Rotation and Robust to Noise (BIRRN) for Floor Recognition

Floor recognition is a conventional task in computer vision with several applications in different fields, from augmented reality to autonomous driving. To address this problem, there is a plethora of methods, several of them based on the use of visual descriptors. However, most previous work has low robustness under image degradation. One alternative to address image degradation problems is the use of binary descriptors. Unfortunately, these descriptors are sensitive to noise. In addition, these descriptors use only some pixels within a patch, this limits the floor recognition scope since useful information is available just for a small pixel set. To cope with these problems, we propose a new texture descriptor based on binary patterns suitable for floor recognition. This descriptor is robust to noise, robust to illumination changes, invariant to rotation and it considers a larger number of pixels than the used in the previous LBP-based approaches. Experimental results are encouraging, the proposed texture descriptor reach high performance under several real-world scenarios, 7.4% more recall and 3.7% F - score than previous texture descriptors and it has high robustness under image degradation.