Sensor-based gesture recognition algorithms are generally divided into the following two types: traditional machine learning and deep learning.

Gesture recognition algorithms based on machine learning (ML) include DTW, support vector machine (SVM), random forest (RF), K-means, and K-nearest neighbors [16,19,20,21]. These methods are widely applicable and adaptable to various types of complex gesture data. At present, many researchers have conducted research on the improvement of related algorithms in sensor-based gesture recognition. Although the ML-based gesture recognition method is relatively simple to implement, the number of parameters generated is also lower than that of neural networks, and the requirements for the computing equipment are relatively low. However, with the increase in gesture types and gesture data sequences, the training data required for learning is also increasing. The accuracy and response time of the recognition algorithm will also be affected to a certain extent.

The basic model of deep learning (DL)-based gesture recognition mainly includes the convolutional neural network (CNN) [22], deep neural network (DNN) [23], and recurrent neural network (RNN) methods [24]. The DL model has become the mainstream classification method in gesture recognition due to its excellent performance, high efficiency in extracting data features, and ability to process sequential data. Fang et al. [25] designed a CNN-based SLRNet network to recognize sign language. This method used an inertial-sensors-based data glove with 36 IMUs to collect a user’s arm and hand motion data, and the accuracy can reach 99.2%. Faisal et al. [26] developed a low-cost data glove deployed with flexible sensors and an IMU, and introduced a spatial projection method that improves upon classic CNN models for gesture recognition. However, the accuracy of this method for static gesture recognition is only 82.19%. Yu et al. [27] used a bidirectional gated recurrent unit (Bi-GRU) network to recognize dynamic gestures, realize real-time recognition on the end side (data glove), and reach a recognition accuracy of 98.4%. The limitation of this approach is that it is not possible to only use the smart glove, but external IMUs must be employed on the user’s arm, which can cause discomfort to the user.

The selected model needs to be determined according to the type of task, requirements, and other factors. Due to the complex amphibious environment, the underwater and land environments are different, and the interference to the sensor is entirely different. It is difficult to transmit Bluetooth signals underwater, and it is difficult to send data to the host wirelessly. Therefore, choosing a gesture recognition model suitable for the amphibious environment is essential. A study addresses this gap by proposing a novel amphibious hierarchical gesture recognition (AHGR) model that adaptively switches classification algorithms according to environmental changes (underwater and land) to ensure recognition accuracy in amphibious scenarios. In addition, it is also challenging to ensure accuracy for cross-user and cross-device recognition using a pretrained DL model. Although some studies on gesture recognition across users and in different environments has made some progress [12], they were mainly focused on EMG-based gesture recognition, and there is a lack of research on cross-user gesture recognition using data gloves based on stretch sensors and IMUs.