Adaptive Edge–Cloud Deep Learning for Efficient Human Activity Recognition

The rapid growth of Internet of Things (IoT) technologies has enabled large-scale sensing and continuous data collection across domains such as healthcare, smart homes, transportation, and wearable systems. In particular, human activity recognition (HAR) using wearable and smartphone sensors has gained significant attention due to its wide range of applications, including health monitoring and context-aware services. However, the high dimensionality and continuous nature of sensor data impose substantial challenges for traditional cloud-centric analytics frameworks. Cloud-based deep learning models offer powerful analytical capabilities but require the transmission of raw or minimally processed data from edge devices to centralized servers. As the number of IoT devices increases, this approach leads to excessive bandwidth usage, increased latency, and potential scalability bottlenecks. These limitations motivate the exploration of alternative architectures that can reduce data transmission while preserving analytical performance. Edge computing addresses this challenge by enabling localized processing closer to data sources. By shifting part of the computation to edge devices, it is possible to reduce network overhead and support real-time analytics. However, edge devices are typically resource-constrained, making the deployment of complex deep learning models challenging. This has led to the adoption of edge–cloud collaborative frameworks, where lightweight processing is performed at the edge and more computationally intensive tasks are handled by the cloud. This work investigates an edge–cloud deep learning framework for human activity recognition that emphasizes deployment efficiency rather than accuracy maximization. Instead of attempting to outperform cloud-only models, the proposed approach focuses on reducing data transmission at the edge while maintaining acceptable analytical performance.