Answers are grounded in this paper and cite their sources.
Monitoring farmer activity can support workload assessment, but wearable activity recognition depends on temporal context and classifier configuration. This study evaluates quaternion-based Sparse Distributed Representations, Hierarchical Temporal Memory (HTM), and supervised readouts using publicly available agricultural material-handling recordings from 20 participants and 186 trials covering six activities. Chest inertial signals were processed at 50 Hz and evaluated at 10 Hz using fixed participant-disjoint training, validation, and test groups across three seeds. The original HTM readout achieved 23.58 ± 0.18% macro-F1 and 54.59 ± 0.71% accuracy. With the HTM backbones fixed, a tuned hybrid-SGD readout achieved 57.41 ± 2.33% macro-F1 and 69.97 ± 1.38% accuracy, while a current-only regularized logistic readout reached 57.79 ± 2.31% macro-F1 and 69.50 ± 1.42% accuracy. Both tuned readouts exceeded the quaternion-only logistic regression baseline (42.20% macro-F1) but remained below CNN (75.67%), LSTM (74.47%), and Transformer (89.29%). The elapsed-time-only baseline achieved 79.86% macro-F1, indicating substantial predictability from the prescribed task sequence. These findings demonstrate strong sensitivity to readout configuration without establishing a consistent benefit from 20 ms auxiliary supervision. Field generalization, noise robustness, and real-time deployment remain unvalidated.
No takes yet. Share an insight, caveat, or question.
Adhitya et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: