A fast monocular 6D pose estimation method for textureless objects based on perceptual hashing and template matching.

Abstract

Object pose estimation is essential for computer vision applications such as quality inspection, robotic bin picking, and warehouse logistics. However, this task often requires expensive equipment such as 3D cameras or Lidar sensors, as well as significant computational resources. Many state-of-the-art methods for 6D pose estimation depend on deep neural networks, which are computationally demanding and require GPUs for real-time performance. Moreover, they usually involve the collection and labeling of large training datasets, which is costly and time-consuming. In this study, we propose a template-based matching algorithm that utilizes a novel perceptual hashing method for binary images, enabling fast and robust pose estimation. This approach allows the automatic preselection of a subset of templates, significantly reducing inference time while maintaining similar accuracy. Our solution runs efficiently on multiple devices without GPU support, offering reduced runtime and high accuracy on cost-effective hardware. We benchmarked our proposed approach on a body-in-white automotive part and a widely used publicly available dataset. Our set of experiments on a synthetically generated dataset reveals a trade-off between accuracy and computation time superior to a previous work on the same automotive-production use case. Additionally, our algorithm efficiently utilizes all CPU cores and includes adjustable parameters for balancing computation time and accuracy, making it suitable for a wide range of applications where hardware cost and power efficiency are critical. For instance, with a rotation step of 10° in the template database, we achieve an average rotation error of $10°$ , matching the template quantization level, and an average translation error of 14% of the object's size, with an average processing time of $0.3s$ per image on a small form-factor NVIDIA AGX Orin device. We also evaluate robustness under partial occlusions (up to 10% occlusion) and noisy inputs (signal-to-noise ratios [SNRs] up to 10 dB), with only minor losses in accuracy. Additionally, we compare our method to state-of-the-art deep learning models on a public dataset. Although our algorithm does not outperform them in absolute accuracy, it provides a more favorable trade-off between accuracy and processing time, which is especially relevant to applications using resource-constrained devices.