Semantic position recognition and visual landmark detection with invariant for human effect

This study was conducted to realize semantic position recognition based on visual landmarks (VLs) used for an autonomous mobile robot in an environment living with humans. This paper presents a novel VL feature extraction and description method that is robust for the effect and interference of humans. The proposed method provides a mask image of human regions using histograms of oriented gradients (HOG). The VL features are described with accelerated KAZE (AKAZE) after extracting conspicuous regions obtained using saliency maps (SMs). We created codebooks as visual words using bags of AKAZE features that have different feature numbers. For scene recognition, we used adaptive category mapping networks (ACMNs) that learn adaptively with additional sequential data and visualization of topological data structures as a category map. As a preliminary experiment, we evaluated the accuracies of human detection using EPFL benchmark datasets. The true positive rate and the false positive rate are, respectively, 85.4% and 41.6% for 100 randomly selected images in four scene categories. As an evaluation experiment, we created our original benchmark datasets using a mobile robot. The recognition accuracy evaluated using leave-one-out cross validation (CV) reveals 49.9% for our method, which is 3.2 percentage points higher than the accuracy of the comparison method without HOG detectors.