Innovative application of neural networks hybrid model based on attention mechanism in natural human-computer gesture interaction

In order to improve the natural interaction ability of quadruped robots, this study proposes a hybrid model (CNN-LSTM-AM) that integrates convolutional neural networks (CNN), long short-term memory networks (LSTM) and attention mechanisms (AM). By improving the infrared sensor data acquisition method (APDS9960 sensor) and optimizing the data set (Net4uCA enhanced set), combined with a deep learning framework, efficient recognition of dynamic gestures is achieved. Experimental results show that the model has an average recognition accuracy of 99% on 1,350 data sets containing 9 types of gestures, which is significantly better than single CNN (92%) and LSTM (88%) models. The innovative introduction of a time step dynamic weight adjustment mechanism effectively captures the global dependencies of time series data and provides technical support for complex human-computer interaction scenarios.