GraDE: A Severity-Graded Dashcam Event Dataset with Per-Video Adaptive Segmentation

The ego-view traffic datasets generally formulate crashes as a binary recognition problem, in which near-misses are largely excluded and collisions are not graded based on their physical severity. GraDE (Graded Dashcam Events) aims to fill this gap by providing a total of 1,321 videos collected from 25 public dashcam compilation uploads. Videos are categorized into three classes: Non-Crash, Near-Miss, and Crash. Among the 487 Crash videos, 433 have a Minor or Major grade indicating their severity as seen in the video footage. A second annotator rated a predefined set of 297 videos; Cohen’s kappa was 0.656 for the 243 videos that neither of the two annotators considered ambiguous. In the compilation videos, unrelated events are spliced together by an editor. To select the clips from compilation videos, we utilized QualityGridAuto, which searches fourteen candidate distance thresholds for each source video and evaluates their quality based on usable duration, boundary strength, within-segment similarity. It applies adaptive thresholding to the construction of the dataset rather than developing a new segmentation method. The frames were sampled at 1 fps and embedded at 224 × 224 pixels, while each baseline utilized sixteen frames with the same resolution. In the manual audit, 93.2%of inspected clips contained one continuous scene. With source-pure four-fold testing, the CNN + BiLSTM baseline reached macro-F1 0.8725 on the three event classes and Video Swin-T 0.9053, but CNN + BiLSTM fell to 0.6889 when the two severity grades were added. On dedicated Minor versus Major classification the two models reached 0.5987 and 0.6510, so severity grading remains substantially harder than event classification. GraDE is released as an annotation index.