Abstract: The gesture recognition system based on YOLOv3 uses darknet53.conv.74 model to train and learn，and separates input images unnecessary information by smoothing and binarizing algorithm，so as to improve the recognition accuracy and realize a video real-time gesture recognition model. The graphical interface is developed by Python Tkinter. The results show that the recognition accuracy of the model can reach 76. 76%， which is comparable with the current mainstream deep learning target detection algorithm. The model is superior to other current mainstream deep learning target detection algorithms in recognition rate，and has advantages in dealing with natural interactive information，which provides an effective means for human-computer interaction.

Key words: deep learning, convolutional neural network, gesture recognition, TensorFlow, YOLO