The confocal unstable cavity has been widely used in chemical laser systems due to its large controllable mode volume, high transverse mode discrimination, near-diffraction limit plane wave output, and high power and high quality far-field spots. The guiding light emitted by the laser is reflected back and forth between the two mirrors of the confocal unstable cavity and interferes with each other. In theory, the beam that output from the confocal unstable cavity is multiple concentric rings. However, due to the influence of air disturbance, platform vibration and other factors, the gray level of the output interference concentric rings fluctuates, the light intensity is gradually weakened from inside to outside, and is in rapid dynamic change. These factors increase the difficulty of extracting the optical pupil position of the guiding light in the automatic optical path collimation. In order to solve the difficult problem of optical pupil position extraction in automatic optical path collimation, a method for optical pupil position extraction of unstable cavity laser is proposed in this paper. First, the original image is preprocessed using the bilateral filtering algorithm. After image preprocessing, calculate the image threshold. Multiplying the coefficient by the image threshold of the Otsu algorithm to get the final threshold of the image. After calculating the threshold, performing threshold segmentation on the image. After the binary image is obtained, the median filter is applied to the binary image to reduce the noise in the binary image and fill the small holes in the binary image. Each contour in the binary image is regarded as a separate connected region, and the contours of the binary image are extracted by using the idea of connected region detection. Then, the inner contour of the optical pupil image can be found from many contours by using the two contour attributes of the contour center and the number of points that make up the contour. After finding the inner contour of the optical pupil image, the Random sample consensus (RANSAC) ellipse fitting algorithm is used to fit the inner contour, which reduces the influence of the internal contour noise on the ellipse fitting results, so as to improve the robustness and accuracy of the optical pupil position extraction results. Since the RANSAC ellipse fitting algorithm needs to fit multiple ellipses, NVIDIA GPU is used to speed up the RANSAC ellipse fitting algorithm, and thousands of ellipses are fitted in the GPU core at the same time to improve the real-time performance of the algorithm. The optical pupil position extraction method proposed in this paper can meet the requirements of real-time, reduce the impact of internal contour noise on ellipse fitting, and improve the accuracy and robustness of optical pupil position extraction results. MATLAB is used to generate 100 frames of simulated optical pupil images with the known position, and then 40% random noise points with amplitude 0~13 are added to the inner contour of the simulated optical pupil image. The proposed method and the direct ellipse fitting method are used to fit the inner contour, to extract the position of the optical pupil image. The simulation results show that the proposed method has higher accuracy and robustness than the direct ellipse fitting method. Compared with the results of direct ellipse fitting, the average value of the extracted position deviation is 0.283 pixels smaller, the maximum value of the position deviation is 0.660 pixels smaller, and the root mean square value of the position deviation is 0.136 pixels smaller.