<![CDATA[Wearable human activity recognition has become a practical foundation for coaching analytics, workload monitoring, and interactive sports training, yet volleyball-specific labelled inertial datasets remain much smaller than general-purpose public HAR corpora. This study addressed that gap through a transfer-learning design in which public HAR benchmarks were treated as representation sources and a smartwatch target domain was used as a volleyball-oriented wrist proxy. Full experimental evaluations were conducted on UCI HAR and on a public WISDM-derived smartwatch subset, using three baseline families: logistic regression, a lightweight one-dimensional convolutional network, and a tiny Transformer. The study also measured modality ablation and unsupervised domain adaptation through Deep CORAL in a common four-class transfer space. On UCI HAR, the final measured accuracies were 0.7940 for logistic regression, 0.8039 for CNN-Lite, and 0.3621 for Transformer-Tiny. On the WISDM smartwatch subset, the corresponding accuracies were 0.8207, 0.8682, and 0.7632. Modality ablation on WISDM showed that accelerometer-only input reached 0.8486 accuracy, gyroscope-only input reached 0.6543, and fused accelerometer-plus-gyroscope input reached 0.8505. For cross-dataset transfer from UCI to WISDM, source-only training achieved 0.3376 accuracy, Deep CORAL improved accuracy to 0.4134, and the fully supervised target-only upper bound reached 0.8374. The results establish three concrete conclusions: lightweight convolutional sequence encoders are more reliable than the tested tiny Transformer under these data conditions, accelerometer channels carry most of the discriminative value for wrist-worn deployment, and domain adaptation is necessary when general smartphone HAR is transferred to smartwatch sports analytics. These findings provide a reproducible public-data foundation for volleyball-oriented wearable modelling and for subsequent fine-tuning on sport-specific action labels.]]>
