The zero-drift compensation of fiber optic gyroscope (FOG) is one of the main methods to improve the working accuracy of FOG. To achieve FOG zero-drift real-time online compensation, an online support vector machine regression (Online-SVR) method based on incremental learning is used to establish the FOG zero-drift real-time compensation scheme. In addition, a real-time temperature change rate acquisition method based on moving average is proposed, which can achieve stable temperature change rate acquisition to meet the requirements of online compensation. The radial basis function neural network, support vector machine regression, and Online-SVR are established by analyzing and preprocessing the measured data of FOG in the range of -15 °C-50 °C. At full temperature, Allan variance is used to analyze the raw and residual zero-drift after compensation of the three models. The results show that the Online-SVR model not only realizes online compensation, but also has better compensation accuracy and stability than the other two models, making it more suitable for online compensation of FOG zero-drift data.