Abstract: In order to solve the problem that BMN (boundary matching network) algorithm only captures the local temporal context when predicting the action boundary, which lead to inaccurate positioning of the boundary in complex scenes, a temporal action proposal generation algorithm based on self-attention is proposed. Specifically, a local-global encoder (LGE) was introduced in the backbone network of BMN to fully mine local and global context. In order to solve the problem that proposal-proposal ignored in BMN, a proposal relation module was designed, which included two self-attention sub-modules. The experimental results on the ActivityNet-1.3 dataset show that AUC improved from 67.10% to 68.20%, an improvement of 1.1 percentage points, reaching advanced performance.