An algorithm for third-party intrusion action detection in oil and gas pipelines based on fine-grained feature enhancement

The extensive distribution of oil and gas pipeline networks across China results in frequent third-party intrusions in high-consequence areas, significantly increasing the risk of pipeline failures and posing serious threats to pipeline safety. Current detection methods mainly rely on manual inspections and video surveillance. However, traditional manual inspections suffer from high workloads, low efficiency, poor effectiveness, and considerable safety risks. Existing video surveillance technologies can only identify abnormal objects within pipeline protection zones, failing to recognize abnormal behaviors effectively. These limitations lead to high false alarm rates and poor recognition capabilities. To address these issues, this study designs a multi-scale network feature extraction structure based on the SlowFast algorithm framework. The design captures fine-grained features of small targets across various scales in complex oil and gas pipeline scenes. The proposed approach enhances spatiotemporal representation by leveraging features across different temporal scales. Corresponding feature fusion methods are also designed for these improvements to develop a third-party intrusion abnormal action recognition technology. This enhances the ability to identify third-party intrusions in oil and gas pipelines and provides support for the intelligent development of pipeline infrastructure.