Binocular vision calibration method for a long-wavelength infrared camera and a visible spectrum camera with different resolutions

We present a calibration plate for the binocular vision system, which is composed of a long-wavelength infrared camera and a visible spectrum camera with different resolutions. The calibration plate mainly consists of a white low-temperature aluminum plate with 7×7 round through-holes, a black high-temperature stainless steel plate, and a heating plate. It can be captured by the long-wavelength infrared camera and visible spectrum camera simultaneously. In order to reduce the influence of thermal crosstalk on the edge and angle sharpness of the thermal image of the chessboard calibration plate, we use the round through-holes to replace the black-white squares in the chessboard calibration plate. Based on the fabricated calibration plate, we also propose a related calibration method. The proposed method can quickly detect the calibration plate by using the YOLO-V4 neural network. The affine transformation is performed to get the front view of the calibration plate, and a novel circular detection strategy based on arc level instead of pixel-level is adopted to detect the edges of the round through-holes in the calibration plate. The centers of round through-holes are detected, and the parameters of the cameras are calculated according to the coordinates of the centers in the image coordinate system. The simulation experiments and error analysis have been done to verify the centers detection method. The simulation results show that the error of center detection is always less than 1.4 (pixel). In order to further verify the performance of the calibration plate and the proposed calibration method, a binocular vision system based on long-wavelength infrared camera and visible spectrum camera is fabricated, and the verification experiments have been done. In experiments, our calibration plate and the proposed method are compared with the famous Zhang’s method. The calibration’s average overall mean errors of the visible spectrum camera and long-wavelength infrared camera are about 0.0126 (pixel) and 0.0238 (pixel), and they are respectively decreased by 78.13% and 81.93% compared with Zhang’s method. The re-projection error of the binocular vision system is about 0.548 (pixel), which is decreased by 24.52% compared with Zhang’s method. The average calibration time of the proposed method is about 0.26s.