6D Pose Estimation Algorithm Based on Improved YOLOv5 With Asymptotic Feature Pyramid Network and Attention Mechanism

Accurate 6D pose estimation is crucial for applications such as augmented reality and robotic arm manipulation. Inaccurate pose estimation can lead to augmented reality failures or robotic grasping errors, directly affecting system performance and safety. However, real-world 6D pose estimation presents significant challenges, including cluttered backgrounds, lighting variations, and the absence of surface textures. These issues make it difficult for existing methods to achieve both high accuracy and real-time performance. To address these challenges, we improve the YOLOv5-based 6D pose estimation algorithm. We enhance the original convolutional neural network model by introducing position-aware channel attention into the backbone network, increasing flexibility and precision in feature processing across different positions. Furthermore, the C3 module is replaced with the lightweight C2F module to improve feature extraction capabilities, maintaining accuracy while reducing model complexity. Additionally, the neck component is enhanced by replacing the BiFPN (Bi-directional Feature Pyramid Network) with the AFPN (Asymptotic Feature Pyramid Network), and the EMA module is integrated into the three C2F modules to boost multi-scale feature fusion and model stability. Finally, the PnP algorithm is applied to calculate the 6D pose. Experiments conducted on the public LINEMOD dataset show that our method significantly outperforms the original YOLOv5-6D method in terms of the 2D reprojection, 5cm5degree, and ADD metrics. The proposed method enhances accuracy and accelerates inference, meeting the requirements for robust pose estimation.