Human Action Recognition Classification Based on 3D CNN Deep Learning

This paper proposes a human action recognition method based on 3D convolutional neural network (C3D), which does not require manual feature extraction and can efficiently and accurately recognize various human actions in complex environments. This paper uses Keras framework to build a C3D model based on InceptionV3, and uses ImageDataGenerator and other tools to enhance and preprocess the data. This paper selects UCF101 dataset as the training and testing data, and uses FFmpeg to convert the videos. This paper evaluates and tests the model, using the test set and custom videos for recognition, and gives the accuracy and top-5 accuracy indicators. The trained model can not only recognize and output the name of the most likely action, but also recognize multiple actions and judge abnormal behaviors.