Human Activity Recognition (HAR) is at the cross-point of pervasive computing, smart home technology, and ambient assisted living, aiming to identify and track everyday behaviors from a continuous stream of sensor readings. This study addresses the challenging task of concurrent human activity recognition, wherein multiple activities may overlap in time or occur in close succession, using a diverse range of ambient sensor data. To tackle this, a Dual Attention-Driven Sync-LSTM architecture is introduced, unifying synchronized Long Short-Term Memory (Sync-LSTM) networks with self-sensor attention and temporal attention. The dual-attention mechanism effectively prioritizes key sensors at each time step and refines focus on the most relevant temporal segments, thereby managing heterogeneous sensor reliability and dynamic data patterns. An overlapping sliding window technique segments the time-series data, ensuring sufficient temporal context is captured. The experiments on real world data like the CASAS smart home corpus, prove that the model is effective in detecting frequent and overlapping activities, especially in the situation of concurrent activities where more than two actions occur simultaneously. The results highlight its ability to capture complex temporal dependencies and sensor interactions, leading to improved recognition of commonly paired activities. Nonetheless, limitations arise in recognizing underrepresented activities, indicating the need for additional strategies such as rebalancing or advanced data augmentation. In general, the results highlight the feasibility of HAR based on attention-based strategies, which has resulted in the interpretability of sensor signals and improved adaptability in the smart living environment. By focusing on sensor diversity and time variations, the suggested model provides a direction on the way to more transparent and context-sensitive HAR systems.

错误:搜索内容不能为空,请输入英文关键词
错误:关键词超出字数限制,请精简
高级检索

Dual Attention-Driven Sync-LSTM for Concurrent Human Activity Recognition

  • S. Jai Harish,
  • K. Nalinadevi

摘要

Human Activity Recognition (HAR) is at the cross-point of pervasive computing, smart home technology, and ambient assisted living, aiming to identify and track everyday behaviors from a continuous stream of sensor readings. This study addresses the challenging task of concurrent human activity recognition, wherein multiple activities may overlap in time or occur in close succession, using a diverse range of ambient sensor data. To tackle this, a Dual Attention-Driven Sync-LSTM architecture is introduced, unifying synchronized Long Short-Term Memory (Sync-LSTM) networks with self-sensor attention and temporal attention. The dual-attention mechanism effectively prioritizes key sensors at each time step and refines focus on the most relevant temporal segments, thereby managing heterogeneous sensor reliability and dynamic data patterns. An overlapping sliding window technique segments the time-series data, ensuring sufficient temporal context is captured. The experiments on real world data like the CASAS smart home corpus, prove that the model is effective in detecting frequent and overlapping activities, especially in the situation of concurrent activities where more than two actions occur simultaneously. The results highlight its ability to capture complex temporal dependencies and sensor interactions, leading to improved recognition of commonly paired activities. Nonetheless, limitations arise in recognizing underrepresented activities, indicating the need for additional strategies such as rebalancing or advanced data augmentation. In general, the results highlight the feasibility of HAR based on attention-based strategies, which has resulted in the interpretability of sensor signals and improved adaptability in the smart living environment. By focusing on sensor diversity and time variations, the suggested model provides a direction on the way to more transparent and context-sensitive HAR systems.