The purpose of action recognition is to identify the action in the video, typically the action of the person within the video. Video can be regarded as a data structure composed of a set of image frames arranged chronologically, having one additional time dimension compared to the image. Action recognition not only requires analyzing the content of each frame image in the video but also needs to extract clues from the temporal information between the video frames. Although deep learning has achieved global success in the field of image classification, its performance in the field of action recognition is not as remarkable as in the field of image classification. For a considerable period of time, the accuracy of action recognition based on deep learning algorithms has been unable to reach or only approach the accuracy of traditional action recognition algorithms. In this paper, a simple model with fewer parameter adjustments is constructed based on 3DResNet-18 and ResNeXt’s split-transform-merge strategy. Action recognition on the KTH dataset has achieved better results than other algorithms, which proves that the improved model has superior performance. The experimental results indicate that the accuracy of the improved algorithm in this paper is 96.3%. Compared with the original 3DResNet-18, the improved model can capture human actions in the original image more effectively and enhance the recognition effect.