The invention discloses a real-time video
stream-oriented edge environment
behavior recognition system, which comprises a video
stream acquisition module, a video
stream storage module, a video stream preprocessing module, a cache module, a sliding window positioning module, a
behavior recognition module and a
user interface module; the sliding window positioning module starts an online detection method based on the action of the elastic
jumping sliding window, a large amount of computing resources are saved by positioning the position of the action, and meanwhile, large loss of the
detection performance of an original model with excellent performance due to the fact that frame-by-frame sliding of the sliding window is abandoned is avoided; a model lightweight effect is realized to a certain extent, a large amount of
invalid data is prevented from being sent to a behavior identification module, the behavior identification efficiency is improved, the identification real-time performance is improved, and the privacy security problem of data acquired by edge equipment in a sensitive scene is better protected through data localization
processing; and compared with a behavior identification service generally based on a cloud center, the
system is more beneficial to the use of edge devices with
limited resources.