A large number of sensors are used in the method of traditional intention recognition. The feature space composed of data has high dimension, and the statistical features are unstable for short-term samples. In this paper, the geometric features represented by the joint angles and the physical features represented by the acceleration and angular velocity are organically integrated, and they are used in the motion intention recognition of the intelligent lower limb prosthesis. Firstly, the knee joint angle is calculated by using the short time series data collected from the healthy thigh and calf in the early swing phase by the inertial measurement unit, so as to obtain the rotation characteristics of the thigh and calf around the joint axis. Then, the mean value and variance of physical features are extracted to reflect the mean level and dispersion degree of short-term data, and the slopes of the maximum and minimum of geometric features are extracted to reflect the local change rate of short-term data and make up for the instability of statistical features. Finally, the geometric features and physical features are fused, and the support vector machine is used to classify 13 daily behaviors. The experimental results show that the recognition rate of 5 kinds of steady states including walking, stair ascent, stair descent, ramp ascent and ramp descent reaches 96.9%. The recognition rate of 8 kinds of transitional states reaches 97.1%. The recognition rate of 13 kinds of states reaches 94.3%. Only the data of two sensors on the healthy side are used to form a 25-dimensional hybrid feature by feature fusion, which achieves rapid dimensionality reduction and reduces the algorithm complexity.