Behavior recognition research based on reinforcement learning for dynamic key feature selection

In the task of behavior recognition based on time-series sequential data, there are often some features that are interference redundancies after feature extraction of the original data by the depth model, and these redundancies will not be beneficial to recognition but will have interference effects. Therefore, it is important to accurately select the features that are beneficial for recognition in behavior recognition tasks. To address the above issues, We propose a reinforcement learning framework, called Dynamic Key Feature Selection Network(DKFSN), aiming to achieve accuracy improvement by continuously exploring the advantages and disadvantages of distinguishing features, eliminating the redundant features that interfere with recognition, and retaining the features rich in quality information. First, feature extraction of the original data using a baseline network to capture depth features and prediction results. Using the depth features as input to a dynamic feature selection network to predict which features are retained and then making a determination to retain key features. Finally, behavior prediction by retained key features and feedback on the selection behavior using a reward function are used for the training of the DKFSN. We validated the validity of DKFSN on two public benchmark datasets.