In traditional recognition methods based on machine learning and manual features, hand-crafted feature extractors and action classifiers based on traditional machine learning algorithms are often used [17]. Action classifiers are used to recognise and classify human movement actions based on the characteristics of that action. For example, Cho et al. [18] used joint distance features for feature extraction. The category of each pose is labelled by an artificial neural network (ANN). Finally, discrete Hidden Markov Models (HMMs) are applied to classify and recognise action sequences. Meanwhile, in order to effectively improve the recognition performance of the system, some researchers have adopted a key-frame-based approach to reduce the processing time of the system [19,20]. A recognition system for human action sequences was developed using traditional machine learning algorithms combined with key-frame selection. In past research, action recognition methods based on traditional machine learning and manual features were combined with great success. However, for the construction and extraction of features [21], they need to rely on human cognition. Moreover, based on human expertise, only superficial features can be learned, making it difficult to cope with the needs of real environments.

In recent years, a number of new methods have been developed, especially regarding the application of deep learning methods in action recognition [22]. The main representative works can be summarized as discussion methods based on convolutional neural networks and discussion methods based on LSTM.

Traditional CNN models are currently limited to processing 2D inputs and are not suitable for the feature capture of 3D skeleton data. To shift CNNs from images to temporal motion sequences, Tran et al. [23] extended traditional CNNs to 3D CNNs, which are more suitable for spatio-temporal feature learning. Related experiments have shown that this scheme outperforms traditional 2D CNNs in terms of analytical functionality. Another common strategy is to employ two-stream CNNs to deal with the problem of capturing motion information between consecutive frames. Zhu et al. [24] proposed a CNN architecture based on a two-stream approach that implicitly captures motion information between adjacent frames and uses an end-to-end CNN approach to learn optical streams. Task-specific motion representations can be obtained while avoiding expensive computation and storage. Since then, many improved models have been proposed, and the two-stream CNN has made significant contributions to the development of motor action recognition [25]. It can even be referenced to realistic and complex real-world environments; for example, Hu et al. [26] introduced a video triple model to obtain additional timestamp information, thus extending behaviour recognition to workflow recognition. Moreover, with extensive simulation experiments, it was shown that the algorithm is robust and efficient in the recognition of real environments.

However, these algorithms have been shown to be only effective for short-term temporal feature learning and are not applicable to long-term temporal feature encoding. With the development of RNNs, LSTM networks suitable for long-term motion sequences have been developed. They have been gradually applied to human action recognition, demonstrating their ability to effectively alleviate the recognition problem of long-term motion sequences [27,28]. Wang et al. [29] introduced long short-term memory (LSTM) to model the high-level temporal features generated by a kinetically pretrained 3D CNN model, with satisfactory results in the recognition and classification of long-term motion sequences. However, the traditional frame-skipping pattern of LSTM [30] also limits performance in action recognition. The problem of data redundancy accompanies the task of the recognition of long-term motion action data.

In recent years, many researchers have turned their attention to joint-aware and attention mechanisms and have achieved good recognition performance in long-term temporal reasoning tasks. Regarding joint-aware-based recognition methods, Oikonomou et al. [31] argue that each action in real life can be effectively perceived by observing only a specific set of joints and associate a specific joint with each action to point out the joint that contributes the most; Shah et al. [32] separately extracted the motion features of each joint using a motion encoder and then performed collective reasoning and selected the most discriminative joint for the recognition task. Regarding recognition methods based on attention mechanisms, Dai et al. [30] proposed an LSTM network based on end-to-end two-stream attention, which can selectively focus on the effective features of the original input image and give different levels of attention to the output of each depth feature map to effectively improve the recognition performance of the model by adopting a visual attention mechanism to address the problem that features of different frames have different learning roles; Li et al. [33] proposed a spatio-temporal attention (STA) network to learn discriminative feature representations of actions by representing useful information at the frame level and channel level, which can be inserted into state-of-the-art 3D CNN architectures for video action detection and recognition with better recognition performance; in the article [34], the authors proposed a bi-directional long short-term memory (BiLSTM)-based attention mechanism. The attention mechanism is used to improve performance and extract additional high-level selective action-related patterns and cues, thereby obtaining a high-performance recognition model.