Evaluation Model of Athletes’ Lower Extremity Training Ability Based on LSTM Algorithm

To achieve intelligent evaluation of the lower limb movement ability of athletes with sports disabilities, this article selects young athletes and middle-aged and young athletes with sports disabilities as the research objects, and healthy young athletes as the control group. Gait videos, GRF and knee angles of the subjects were collected to extract and analyse gait contours and features. The improved visual background extractor (ViBe) algorithm has the highest accuracy of 0.978 in PETS2006 video sequences; the recall rate of three algorithms in Highway video sequence is the highest, and the recall rate of improved ViBe algorithm is the highest, up to 0.965. At Ɛ=0.7,, the accuracy of the improved myloss training set is higher than that corresponding to other values; when the number of iterations is 98, the accuracy rate of improved myloss training set is 0.963, while the accuracy rate corresponding to the cross entropy loss function is 0.945. When the number of iterations is 151, the accuracy rate of the Xeption LSTM model is 0.956, higher than that of other models. Among corresponding mean ± standard deviation (GSA-MS) indicators, the GSA-MS values of Group L are significantly higher than those of Group N and Group Z (P < 0.001). The correlation between GAS indicators and Gait Abnormality Rating Scale (GARS-M) is strong, with a correlation coefficient of 0.90.