Aiming at the deficiency of similar motion recognition in the current UWB radar attitude recognition research domain, a motion recognition model integrating time-frequency analysis and random forest (RF) is proposed. A time-frequency analysis method of two-dimensional smoothed pseudo Wigner-Ville distribution (2D-SPWVD) based on smoothed pseudo Wigner-Ville distribution (SPWVD) is proposed to extract the time-frequency features of the preprocessed human motion echo signals. Principal component analysis (PCA) was employed to reduce the dimension of the feature vectors, and the top 30 principal components with a high cumulative contribution rate were selected as new feature vectors to be input into the RF classification model optimized by sparrow search algorithm (SSA) for the identification of five distinct human similar drop actions in the presence of obstacles. The experimental outcomes demonstrate that the pretreatment algorithm can effectively enhance the SNR of the action echo signal, and the PCA-SSA-RF classification model can effectively distinguish five different human fall movements, overcome the particularity of data and the interference of obstacles, with an accuracy rate as high as 96.6%. In the fall detection task within the real-time data stream, the average classification accuracy of the model reaches 93%, and it is profoundly compared with RF, PSO-RF and other diverse classical classification models, featuring high accuracy and short overall time, and possessing both accuracy and classification efficiency. The superiority and effectiveness of the proposed method are verified.