In this paper, a multi-parameter fusion neural network (FNN) for human actions recognition based on multi-antenna frequency modulated continuous wave (FMCW) radar is proposed. First, in view of the shortcomings existing in FMCW radar parameter estimation algorithms, we develop a joint range-azimuth parameter estimation method combining minimum power distortionless response (MPDR) beamforming with fast Fourier transform (FFT), which provides higher angular resolution and overcomes the problem of the performance degradation that most of the super-resolution algorithms tend to have under unknown target number. Second, this method uses two mutually perpendicular linear array radars to capture human actions, and then the distance as well as angle parameters of the signal reflection position of human target in the horizontal and vertical directions are estimated. After that, FNN is constructed to extract and fuse the spatial and temporal features of human action in the horizontal and vertical directions from the parameter estimation results, based on which this method realizes human actions recognition and classification. The extensive experimental results show that FNN proposed improves the recognition accuracy by 4.37% compared with the traditional 3D convolutional neural networks (3D-CNN).