A method and an apparatus for detecting a scene change frame in a video and a system, where when a scene change frame in the video is detected, whether a maximum P frame (Pmax) in all P frames in a group of pictures (GOP) of the video is a scene change frame is determined based on a relative relationship between a size Pkmax of Pmax and a median or an average of sizes of a plurality of I frames between Pmax and a scene change frame closest to Pmax and before Pmax, or a relative relationship between Pkmax and a median or an average of sizes of a plurality of P frames in the GOP to detect a scene change frame in the P frames such that missed detection of a scene change frame is effectively reduced.