Machine vision displacement measurement technology provides a new solution for linear control of large-span bridges, and ensuring high-precision two-dimensional to three-dimensional coordinate conversion is crucial. A method based on improved genetic algorithm BP neural network is proposed to improve the calibration accuracy of binocular cameras. By improving the crossover and mutation probability functions in traditional neural networks, the calibration efficiency and accuracy are improved. Through corresponding experimental examples, it has been verified that the mean square error of mea-suring coordinates using the traditional Zhang calibration method is 4.67 mm. After applying this method for calibration, the mean square error of measuring coordinates is 0.82 mm, which improves the calibration accuracy and can meet the monitoring requirements of bridge construction linearity.