A deep learning system that learns a discriminative model autonomously using difference images.

In our proposed method, an object can be detected from time-series images taken by two or a few cameras. When one of the cameras detects that the object has moved, a system locates that object from the images taken by the other camera(s) by using the timestamps. Then, a position data of that object can be obtained autonomously without taking a lot of photographs of that object and teaching data to the system in advance. Moreover, the system passes the position data and a label of that object to YOLO, which is a learning model for discriminating objects. YOLO learns the data and become possible to indicate the label of the object. The results of an experiment showed that this system could discriminate the moving object only by two cameras without the previously prepared teaching data.1