The Role of Compression Models in Human Activity Recognition: A Comprehensive Review

Human activity recognition has gained significant interest in a number of fields, such as wearable technologies, smart environments, and healthcare. However, the growing complexity of HAR models and the large amount of sensor and video data pose some issues related to storage, computation, and real-time processing. To address these challenges, compression techniques have been explored to reduce model size while persevering accuracy of recognition. This paper compares different compression techniques based on key performance metrics, including compression ratio, accuracy retention, inference time, and energy efficiency. In addition, we discuss how they affect real-time HAR applications, especially in contexts with limited resources and edge computing. Our research demonstrates the trade-offs between performance and compression, the applicability of different approaches for distinct HAR situations, and the challenges associated with computational overhead and information loss. In order to increase the effectiveness of HAR models, we highlight the need for adaptive and hybrid compression techniques in our final research directions. The goal of this review is to provide insight in creating HAR systems that are more efficient and scalable.