Most online action detection methods focus on solving a $K+1$ classification problem, where the additional category represents the ‘background’ class. However, training the ‘background’ class and dealing with data imbalance during training are common challenges in online action detection. $T_{0}$ address these challenges, we propose an effective model to mitigate the negative effects by incorporating an additional pathway between the feature extractor and action identification model Furthermore, we present two configurations for retaining the feature distinctions and supporting the final decision of the Long Short-Term Transformer (LSTR), aiming to enhance its performance in the $K+1$ classification. Experimental results show that the proposed method achieves an accuracy of 71% in mean average precision (mAP) on the Thumos 14 dataset, outperforming the 69.5% achieved by the original LSTR method.