Researchers at Vellore Institute of Technology presented DashGuard, a system that uses dashcam video to anticipate road accidents. In tests on the Car Crash Dataset, its Two-Stream Transformer model achieved an AUC of 0.981, an F1 score of 0.870 and recall of 0.991. The dataset contains 1,500 crash videos and 2,000 normal videos. The study describes dashcam-based anticipation as a low-cost complement to sensor-heavy advanced driver-assistance systems, noting that delayed hazard perception is implicated in most collisions and road accidents claim roughly 1.35 million lives each year.

The researchers trained and evaluated six spatio-temporal models: a 3D CNN, CNN+LSTM, CNN+Transformer, and three two-stream variants using a lightweight CNN, ResNet-18 or Transformer. They used a video-level 70/15/15 split to keep clips from the same video from crossing between training, validation and test sets. Four-level soft labels represented how far each clip was from an accident frame, while a two-clip temporal consistency gate reduced false-positive alerts.

The lightweight 3D CNN scored an AUC of 0.949 and has approximately 180 K parameters, which the researchers describe as a viable option for edge hardware. They report that separating motion and appearance through two-stream attention produced the clearest gains over single-stream designs on this task. A web interface built with React and FastAPI supports interactive evaluation. The research received no funding, and the authors declared no competing interests.

Researchers at Vellore Institute of Technology presented DashGuard, a system that uses dashcam video to anticipate road accidents. In tests on the Car Crash Dataset, its Two-Stream Transformer model achieved an AUC of 0.981, an F1 score of 0.870 and recall of 0.991. The dataset contains 1,500 crash videos and 2,000 normal videos. The study describes dashcam-based anticipation as a low-cost complement to sensor-heavy advanced driver-assistance systems, noting that delayed hazard perception is implicated in most collisions and road accidents claim roughly 1.35 million lives each year.

The researchers trained and evaluated six spatio-temporal models: a 3D CNN, CNN+LSTM, CNN+Transformer, and three two-stream variants using a lightweight CNN, ResNet-18 or Transformer. They used a video-level 70/15/15 split to keep clips from the same video from crossing between training, validation and test sets. Four-level soft labels represented how far each clip was from an accident frame, while a two-clip temporal consistency gate reduced false-positive alerts.

The lightweight 3D CNN scored an AUC of 0.949 and has approximately 180 K parameters, which the researchers describe as a viable option for edge hardware. They report that separating motion and appearance through two-stream attention produced the clearest gains over single-stream designs on this task. A web interface built with React and FastAPI supports interactive evaluation. The research received no funding, and the authors declared no competing interests.