A Feature Fusion Method Integrating Gradient Boosted Feature Selection and Deep Learning for Performance-Aware Physical Activity Recognition

Recognition of human physical activities (HAR) is a widely used research field due to its prominent applications, such as smart health, housing care, and lifestyle and behavior monitoring. Recently HAR is also attracting much attention in the field of Industry 4.0 as a building block for monitoring workers' activities and improving safety in the workplace. However, assuring high accuracy and low training time of recognition is a challenging problem that slows down the development of smart applications and requires more investigation. In this paper, we propose a solution that combines the gradient boosted feature selection (GBFS) method and convolutional neural network to improve the performance of the HAR model using a smart device. The GBFS is used to select the most relevant and non-redundant features from the frequency and time domain by enhancing the performance of the proposed model. While most deep learning methods use fine-tuned hyperparameters to improve recognition accuracy, in our approach, enhancement of the recognition model is reached by a suitable architecture, and as a consequence, we also reduce the timing of training. The UCI public dataset and a self-own dataset are used to assess the generalization capability and performance of the proposed feature fusion method.