Abstract: To solve the problem that the comprehensive utilization of overall-local-aware spatio-temporal relationship information was not considered in current action prediction and localization and was not conducive to improving the performance of action detection and localization, a temporal action detection method based on overall-local-aware graph network was proposed. To obtain richer overall spatio-temporal feature representation of proposals, the feature similarity and temporal overlap of each action proposal was comprehensively exploited to construct the overall relation graph reasoning sub-network of proposals. To obtain local relation information of proposals under different time scales, the partial order relationship over time for the proposals was exploited, and the local relation graph reasoning sub-network was constructed, which consisted of multiple levels of three-body similar graphs and three-body complementary graphs. The rich overall-local aware features for the proposals were represented, which were used to predict and localize actions. The experiments were conducted on two public datasets of Thumos14 and ActivityNet1.3 and measured by the mean average precision metric. The results show that compared with the advanced methods of PGCN, G-TAD, TAL-Net and CDC, the proposed method can effectively improve the performance of action detection.