In order to monitor the behavior of rough handling of parcels in the process of express sorting in real time, a visual recognition system of human behavior in violent express sorting based on Raspberry Pi + Edge TPU was designed. Based on TensorFlow deep learning framework, PoseNet model is used to collect human posture data in real time, LSTM + Attention model is used to realize human action recognition, and Mobile SSD is combined to perform scene recognition so as to realize human behavior visual recognition of violent express sorting. Experiments show that the visual recognition method proposed in this paper can realize the fast and accurate recognition of five kinds of violent sorting activity, and the test accuracy of LSTM + Attention human action classification model reaches 80%. The embedded violent sorting behavior recognition system based on this method can monitor the behavior of rough handling parcels in express sorting in real-time, and give real-time warning.