Title: LastAct: Trajectory-Guided Latest-Activity Localization for Real-Time Smart-Home Activity Recognition

Abstract:

Human Activity Recognition (HAR) powered by ambient sensors is foundational to smart-home systems, particularly for health monitoring and assisted living scenarios. However, practical implementations face significant challenges: sensor data arrives as a continuous stream, and the precise boundaries between activities are rarely known in advance. Consequently, sliding-window inference often generates segments that span multiple activity transitions, resulting in "mixed" windows. This phenomenon creates boundary contamination, which undermines the pre-segmented instance assumption relied upon by most existing benchmarks and models. Additionally, many current pipelines fail to fully leverage spatial context, often treating sensor identifiers as isolated tokens rather than interconnected spatial elements.

To address these issues, we introduce LastAct, a trajectory-centric framework designed for streaming HAR in smart homes. LastAct specifically targets the identification of the most recent activity within mixed windows while explicitly accounting for spatial structures. The framework projects sensor events onto the home’s floorplan, generating a sequence of layout-aligned trajectory images that maintain spatial continuity. To handle contamination, a lightweight gate detects affected windows, while a boundary localizer pinpoints the most recent transition. This enables boundary-guided masking, a technique that highlights evidence occurring after the boundary while suppressing outdated contextual information. For computational efficiency, the system utilizes a precomputed layout-aligned template cache, eliminating the need for repeated rendering.

Empirical evaluations across four public smart-home datasets, conducted under near-realistic mixed-activity protocols, demonstrate LastAct’s effectiveness. The model achieves competitive or superior performance on pure activity windows and delivers substantial gains in Macro-F1 scores for cross/mixed windows. These results highlight LastAct’s enhanced robustness in realistic sliding-window environments.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC