Approach for Smartwatch-Based Complex Human Activity Recognition in Internal Logistics Systems

As the domain evolves, studies in Human Activity Recognition (HAR) are pivoting from the classification of simple, singular actions to the real-time identification of complex, multi-step activities. This study proposes an approach for real-time multi-step activity recognition, specifically demonstrating its application in internal logistics systems based on autonomous guided vehicles through intelligent human activity analysis. The approach leverages such techniques as ensemble learning with classifier stacking, transfer learning, and continuous wavelet transform (CWT). To support the aim of this study, a smartwatch-based data gathering system was developed. The data were collected from human staff representatives engaged in the continuous execution of one of two predefined aggregates that were decomposed into four basic activities. A scalogram-pre-trained DenseNet121 model was utilized as the base classifier, achieving an F1 score of 91.01% for basic activity classification. For complex activities, a convolutional neural network was used as the meta-classifier, achieving an F1 score of 87.44% on the best-performing dataset split strategy. The resulting performance, considering the challenges faced and the method's adaptability, indicates the considerable potential of the proposed method. The proposed approach can be integrated to create a robust framework for HAR, providing a positive impact on intelligent enterprise management.