The petroleum and petrochemical industry involves high-risk operations, where traditional manual supervision suffers from blind spots and incomplete coverage, while massive video data remain underutilized. This study collected 1.4 million images from high-risk operation sites and constructed a multi-mechanism hazard identification method using computer vision, integrating object detection, pose estimation, and object tracking. Spatiotemporal attention mechanisms were incorporated to enhance recognition accuracy for multi-scale and small targets. Based on violation behaviors, an algorithmic reasoning logic was designed to automatically identify key targets from complex video images. The study developed 40 video recognition algorithms for operational hazards (e.g., personnel standing under a crane boom and working at heights without a safety harness), achieving an accuracy of ≥90%. These algorithms enable real-time, intelligent identification of violation behaviors, facilitating the transformation of risk management from “human-based defense” to an integrated “human + technical + intelligent defense” model, allowing early intervention and elevating safety risk management standards.