GenHAR: Generalizing Cross-domain Human Activity Recognition for Last-mile Delivery
Researchers have developed GenHAR, a new framework to improve human activity recognition (HAR) by addressing domain shifts in sensor data. GenHAR learns domain-invariant representations by tokenizing sensor data and analyzing correlations across dimensions, enhancing model robustness. The framework also incorporates selective masking and an efficient attention mechanism to boost performance and reduce computational load. In real-world tests, GenHAR achieved a 9.97% accuracy improvement over existing methods and was deployed to detect over 2.15 billion activities across four cities. AI
IMPACT Enhances the accuracy and efficiency of human activity recognition systems, potentially improving applications in logistics and other sensor-based monitoring fields.