Distortion Correction for Wide-Angle Videos of Power Inspection Robots

To address the issue of distortion in video images captured by wide-angle cameras on power inspection robots, we propose a camera self-calibration algorithm based on image feature matching. First, a Transformer model is used to extract local features from two consecutive video frames, and position encoding is applied. Feature descriptors for both images are obtained using self-attention and cross-attention layers, enabling feature point matching from coarse to fine. Based on this, the displacement of matched feature points in both distorted and undistorted regions is calculated. Using a single-parameter division model, the distortion parameters for different positions in the image are then determined. Finally, the average segment length in the corrected images is used to evaluate the distortion parameters, and the optimal distortion parameter is selected. This algorithm does not require complex geometric features from the scene, making it widely applicable. Experimental results demonstrate that the proposed algorithm achieves effective correction of distorted video images with high solution accuracy.