Different human posture may express different current states and needs of human body. At present, human posture recognition can be roughly divided into visual technology and sensor technology. However, the image extracted by the camera is susceptible to illumination and background interference, and the recognition accuracy and robustness are not ideal under complex conditions. Traditional sensors (capacitive, voltage, etc.) are vulnerable to electromagnetic interference, so they are not suitable for human posture recognition applications. In recent years, Fiber Bragg Grating (FBG) sensors have been widely used in structural monitoring because of their light weight, anti-electromagnetic interference, wavelength division multiplexing and other advantages. In order to further improve the recognition accuracy of human posture in bio-medicine and kinematics, this paper designs a smart insole based on FBG sensor for human posture recognition, and combines K-fold cross-validation support vector regression algorithm to improve the recognition accuracy. COMSOL simulation software is used to analyze the force on the sole of the foot, and the four main force parts of the big toe, the first metatarsal bone, the third metatarsal bone and the heel are determined. Using the method of wavelength division multiplexing, four FBG sensors with different center wavelengths are arranged in these four positions, and PDMS material packaging is used to design smart insoles. And each FBG sensor is need to be calibrated, the experimental results show that the FBG sensor can maintain a stable measurement performance during the pressure process at constant room temperature, and the fitting linear relationship indicates that the change of its center wavelength is proportional to the load, and verifies that the wavelength offset of FBG has a high linearity and sensitivity to the pressure. In this work, a total of 25 participants are recruited to accomplish eight human postures, including standing, sitting, standing on one foot, folding forward, leaning forward, leaning back, half squat and full squat, respectively. The changes in the center wavelength of the string FBG sensor are recorded. In order to reduce errors, each posture is repeated twice and the average value of its center wavelength offset is taken to construct the data set. The K-fold cross-validation Support Vector Regression (KCV-SVR) model is introduced for data processing, meanwhile, the K value is set to 5, wavelength offset is taken as input, and different postures are taken as output. After K-fold cross-validation, the optimal value of SVR penalty factor and radial basis function parameter are obtained automatically, which are 0.5 and 8 respectively. The experimental results show that the Root Mean Square Error (RMSE) of SVR regression model is 0.510 6, the Mean Absolute Error (MAE) is 0.132 3, and the coefficient of determination R² is 0.967 7. RMSE, MAE and R² of KCV-SVR regression model decreased by 0.460 4, 66.3% and 3.3%, respectively. By comparing the prediction error of SVR and KCV-SVR regression model, the maximum error of SVR is 0.309 6, the minimum error is 0.012, and the average error is 0.088 4. However, the maximum error of KCV-SVR is 0.301 6, the minimum error is 0.001 5, and the average error is 0.057 6. It can be seen that the prediction result of KCV-SVR regression model is better than that of SVR regression model. KCV-SVR regression model basically realizes the effective recognition of human postures, and provides a new idea for the recognition of human postures based on flexible FBG sensor.