This article proposes a movement intervention strategy for young children’s bad posture based on video surveillance and torso recognition algorithms, which solves the situation of similar but not identical actions of different people in monitoring, effectively solves the problem of non-fixed speed behavior classification, and can better Perform behavioral classification tasks. Before children's gymnastics intervention, the control group (10.12 ± 2.97) and the experimental group (8.97 ± 2.72) had no significant difference in gross movement scores (F=0.933, P>0.05) ; After the intervention of children's gymnastics exercise, the control group (12.69 ± 3.08) and the experimental group (15.32 ± 3.11) have higher gross movement scores than before the intervention (t values are -2.346, -2, respectively) 021, P values are all <0.05).