A coaxial long-short-focus camera-based vision tracking and localization system for high-precision measurement

To address the limitations of traditional binocular vision systems—such as low pixel resolution, short measurement range, and poor adaptability—this paper proposes an integrated framework for multi-stage visual enhancement and high-precision tracking and localization based on a co-axial fusion of long- and short-focal-length cameras. The hardware utilizes a coaxial setup where the long-focal camera provides high-resolution detail for distant small targets—overcoming the resolution bottleneck of the short-focal camera. By precisely calibrating the optical center coordinate of the long-focal camera, we eliminate the complex challenges associated with stereo calibration and feature matching across different focal planes, ensuring highly accurate field-of-view overlap. In the algorithmic phase, the framework introduces an Image Restoration and Detail Enhancement method utilizing Adaptive Gamma Correction, a Dehazing sub-network, and the Swin Transformer Image Restoration (SwinIR) network. This significantly improves the system's resilience to low-quality visual input under low-light and hazy conditions. For tracking and localization, the system integrates YOLOv11 with the Extended Kalman Filter (EKF) for robust detection and high-accuracy trajectory prediction. Furthermore, a Real-Time Error Correction Mechanism (RTECM), based on turntable and camera feedback, is implemented to model and correct dynamic positioning deviations caused by gimbal control errors and calibration drift, thereby enhancing system robustness. The system employs an electric turntable for pose adjustment using spherical coordinate compensation, ensuring the target remains centered within the co-axial views. Final 3D coordinates are calculated using the forward intersection principle constrained by the co-axial configuration. Experimental results demonstrate that the proposed method enables accurate detection and tracking of moving targets while maintaining alignment within the shared field of view. The system achieves an average localization error of less than 10 cm within 10 m, and less than 30 cm within 25 m, indicating high localization accuracy in short- to medium-range scenarios.