DashGuard: spatio-temporal deep learning for dashcam-based road accident anticipation

Road traffic accidents claim roughly 1.35 million lives each year, with delayed hazard perception rather than mechanical failure implicated in most collisions. Dashcam-based accident anticipation offers a low-cost complement to sensor-heavy ADAS, yet controlled comparisons across architectures remain rare. We present DashGuard, which trains and evaluates six spatio-temporal models a 3D CNN, CNN+LSTM, CNN+Transformer, and three two-stream variants (lightweight CNN, ResNet-18, and Transformer) on the Car Crash Dataset (1500 crash and 2000 normal videos) under a consistent, leak-free protocol. A video-level 70/15/15 split prevents clip-level contamination; four-level soft labels encode temporal distance to the accident frame; and a two-clip temporal consistency gate substantially cuts false-positive alerts. On the held-out test set, the Two-Stream Transformer achieves AUC = 0.981 and F1 = 0.870, with recall of 0.991. The lightweight 3D CNN (AUC = 0.949, \(≈ \) 180 K parameters) is a viable option for edge hardware. A React/FastAPI web interface supports interactive evaluation. The results show that explicitly separating motion and appearance via two-stream attention gives the clearest gains over single-stream designs on this task.