The existing datasets used in video online detection research are mainly concentrated on long videos and the category is relatively simple. At the same time, the detection and evaluation system that meets the needs of online setting is needed to meet the growing demand for short video applications on mobile phones. This paper proposes a new task called online highlight start detection (OHSD) in short video scenarios to assist in guiding the mobile phone to automatically capture highlights or other short video applications during the shooting process. The specific experiment is as below: (1) construct a short video dataset called Highlight45, which is carefully temporally labeled based on mobile phone shooting, to fill the gaps in training and evaluation data for new tasks; (2) set the online evaluation metric——the average precision of the first detection, and the visualization results show that this metric is more suitable for the online start detection task requirements; (3) design a hybrid dual-stream network with sequence contrastive loss function as the baseline method for this task. The experiments show that, compared with the traditional method, the proposed method has achieved 6.98% and 4.11% performance improvements in the existing start detection metric and the average precision of the first detection respectively.