Recognizing human actions in videos quickly and effectively，has broad application prospects and potential economic value. Deep learning has been widely used for action recognition. We proposed self-attention temporal segment networks，whose inputs are clipped video clips. This network is based on deep networks and non-local means. By adding non-local modules to temporal segment networks with ResNet as the basic model，we can get our new model. Verified on TDAP dataset，our new model can recognize human actions more accurately than the original model，without increasing much time complexity.