Evidential reasoning-based action segmentation fusion model for industrial behavior procedures monitoring

초록

Industrial behavior procedure monitoring is critical for ensuring industrial safety. Temporal action segmentation (TAS) provides a security guarantee by analyzing video frames to identify behavior procedure standardization and detect anomalies. However, existing TAS methods face challenges like redundant behaviors and action similarity in industrial scenarios. These factors cause high temporal uncertainty, leading to over-segmentation and boundary ambiguity. To address these issues, this study introduces Evidential Reasoning (ER) rule and proposes an ER-based Action Segmentation Fusion (ER-AS) Model. This model decomposes industrial behavior monitoring into two complementary perspectives: State Identification and Category Identification, with their outputs meticulously integrated through ER decision-level fusion to mitigate uncertainty systematically. Specifically, we propose two novel models, Multi-stage Bidirectional Causal Convolutional Network for State Identification and Multi-grained Branch Transformer for Category Identification. These models enable targeted prediction and enhanced computational efficiency. Furthermore, a multi-stage ER fusion approach is proposed. This method efficiently integrates multi granularity category identification results and state identification results complementarily through two-stage fusion. Comprehensive experiments across multiple metrics demonstrate that ER-AS achieves superior performance while significantly lowering computational complexity.

키워드