Robust Object Recognition Under Partial Occlusions Using an RGB-D Camera

For a robot to execute a specific task, the robot firstly has to recognize what objects are in robot’s view. To complete a specific task in a given time, the computation time for recognition is also important. There are much research for increasing recognition accuracy, but the recognition speed is not enough to be applied in real environment. On the other hand, there are also much research for reducing the computation time for recognition, but the recognition accuracy needs to be further improved. Nowadays, deep network has come into the spotlight due to its speed and accuracy. Deep network doesn’t need to find hand-tuned features. This paper proseses a deep network-based object recognition algorithm. The main contribution is that objects could be recognized under occlusion, as objects are often laid to overlap each other. The occlusion makes object recognition accuracy worse. To overcome this problem, the dataset for training consists of not full images but partial information of images and corresponding ground truths. The object region could be found very quickly by using an RGB-D camera. By assuming that most objects are on the stable plane, object regions are taken easily. Experimental results demonstrate such consideration of contextual information (e.g. objects are on the table) makes the performance of recognition better.