DRVEM-Net: A deep residual volume error modeling network for stereo vision system

Stereo vision systems often exhibit residual systematic errors even after rigorous geometric calibration, a limitation that is unacceptable in high-precision measurement and perception applications. To address this, volumetric error modeling (VEM) establishes a comprehensive error distribution model throughout the 3D measurement space, enabling further data compensation. However, existing VEM techniques are constrained by their reliance on expensive equipment, such as laser trackers or coordinate measuring machines (CMMs), for acquiring ground-truth data, rendering the process inefficient and cost-prohibitive. Furthermore, the task of predicting the entire error field from sparse data is inherently challenging due to the complex, highly non-linear nature of the error sources, for which conventional model-based and traditional neural network approaches often fail to achieve an accurate fit. To overcome these limitations, this paper proposes a novel deep residual network to address this highly nonlinear fitting problem and model the spatial error distribution. The proposed network leverages an improved positional encoding scheme to improve the representation of the spatial error distribution and employs a multilayer funnel-shaped deep residual backbone to enhance its high-dimensional, nonlinear fitting capabilities. Currently, we introduce a calibration target-based methodology to acquire the ground truth data required for network training. Experimental results demonstrate that the proposed VEM method achieves superior modeling accuracy compared to current state-of-the-art approaches. In addition, the proposed calibration target-based data acquisition strategy proves to be more practical and accessible to correct errors in a real-world stereo vision system.