Object detection for robotic grasping using a cascade of convolutional networks

Robot guidance in industry is a significant issue that needs to be dealt with in modern manufacturing facilities. One of the common tasks in this area is the pick and place problem. For proper implementation of an automatic pick and place application using a robotic arm for object grasping, it is necessary to detect the accurate pose of the objects of interest. In this contribution, a novel engineering approach to object positioning, based on image processing is proposed. In this approach, the operation is composed of a cascade of convolutional neural networks. This cascade consists of 2 different types of networks. The first one is the object detection network called YOLOv5. It is used to process the raw image data from the scene to provide precise localization and determine the position of the objects of interest. After that, crops of the detected objects are created and processed by the second neural network, namely EfficientNet. This classification network is used to determine the rotation angle of the detected objects. The proposed approach provides a precision rate of 0.997 and a recall rate of 0.999 for locating and determining the correct position. For angle classification, EfficientNet provides an accuracy of 0.951. All tests are performed on the testing set of the legitimate positioning problem.