Target Detection and Robotic Arm Grasp Pose Estimation Based on YOLOv5 and Transfer Learning

Abstract

For traditional machine learning algorithms, visual recognition algorithms have low recognition accuracy and slow running time. This research studies the scene of the robot doing housework in the family scene, and uses the RGB image information as input to complete the grasping pose estimation of the target object. Based on the object detection model YOLOv5s, the network architecture is built by combining data enhancement and transfer learning with its advantages of lightweight and fast speed. After building a family scene data set to enhance the data of a small number of training samples, the model is trained on the target data set using transfer learning, and the parameters are fine-tuned at the same time. The positioning information of the target object is transformed into the grasping pose of the robotic arm through coordinate transformation, and the robotic arm is controlled to finally complete the grasping task with a fixed grasping posture. Finally, the effectiveness of the algorithm is verified by building an experimental platform and manipulating the UR5 robotic arm to carry out actual grasping experiments. The proposed method based on target detection is fast, has high real-time performance, and has a false/missed recognition rate of less than 2%. The application in the robotic arm can efficiently complete the task.

Keywords

• YOLOv5s;Robotic arm;Pose estimation;Target detection;Transfer learning