Identification of original markers by object recognition algorithms using machine learning

We have been working on object detection by combining markers we design and machine learning. Object detection using such markers is particularly powerful under poor conditions where it is difficult to recognize the detection target such as the influence of occlusion is significant. In our previous study, we proposed markers which are robust to occlusion. In that study, we used SSD to identify markers with an accuracy of approximately 0.85. In present study, we propose two types of markers. One is an improvement of the previously proposed markers. It outperforms previous markers' accuracy. The other uses a reflective material and allows detection from images captured by infrared cameras. For this reflective marker, we use two-step detection using CNN to identify. It increases the situations in which the markers can be used.