STV-FSANet: Track-Level Spatio-Temporal Verification for Fire and Smoke Alarm Validation in Video Surveillance

Area of Science:

• Computer Vision
• Artificial Intelligence
• Surveillance Systems

Background:

• Real-world video surveillance faces challenges in reliably detecting fire/smoke due to visual ambiguities like sunlight and steam, leading to false alarms.
• Existing methods struggle with unstable frame-level detections, necessitating improved validation techniques.

Purpose of the Study:

• To develop a robust Spatio-Temporal Verification Network for Fire and Smoke Alarm Validation (STV-FSANet) to accurately identify fire and smoke events.
• To reduce false alarms in video-based fire detection systems by incorporating temporal verification.

Main Methods:

• A detect-track-verify framework was implemented, localizing candidate regions, associating them into tracks, and verifying sequences online.
• The verifier integrates appearance features with geometric cues and uses a dual-branch GRU for temporal modeling.
• A new dataset, FSAV Tracklet Dataset, was created with over 2 million annotated frames for training and validation.

Main Results:

• The STV-FSANet achieved 97.09% test accuracy and 95.81% macro-F1 score.
• The system demonstrated effective real-time performance, with compact temporal evidence proving more valuable than long histories.
• High recall (90.2%) was achieved in rejecting false fire/smoke tracklets, validating the hard-negative modeling approach.

Conclusions:

• The STV-FSANet provides a reliable and accurate solution for early fire/smoke detection in video surveillance.
• The decoupled design allows for future integration with advanced detector backbones.
• Efficient deployment on GPUs and edge devices was demonstrated, enabling practical real-time applications.