We describe a zero-shot pipeline developed for the ACCIDENT @ CVPR 2026 challenge. The challenge requires predicting when, where, and what type of traffic accident occurs in surveillance video, without labeled real-world training data. Our method separates the problem into three independent modules. The first module localizes the collision in time by running peak detection on z-score normalized frame-difference signals. The second module finds the impact location by computing the weighted centroid of cumulative dense optical flow magnitude maps using the Farneback algorithm. The third module classifies collision type by measuring cosine similarity between CLIP image embeddings of frames near the detected peak and text embeddings built from multi-prompt natural language descriptions of each collision category. No domain-specific fine-tuning is involved; the pipeline processes each video using only pre-trained model weights. Our implementation is publicly available as a Kaggle notebook.
Figures & tables
Figure 1: Chronological sampled frames from a synthetic CARLA traffic incident in the ACCIDENT @ CVPR 2026 dataset. Most clips record approximately 18 seconds of motion at 20 FPS resolution.
Figure 2: Temporal localization on a synthetic CARLA video. Top: mean absolute frame difference dt across all frames. Bottom: z-score anomaly series zt after rolling-mean smoothing. The dashed line marks the detection threshold.
Figure 3: Cumulative Farneback optical flow magnitude M(u,v) on a synthetic CARLA video after 90th-percentile thresholding. The bright region corresponds to the collision area; diffuse background motion is suppressed.
Type
Example prompt
head-on
“two cars colliding head-on from opposite directions”
rear-end
“a car colliding into the back of another car”
sideswipe
“two vehicles scraping alongside each other”
single
“a single car crashing into a wall or obstacle”
t-bone
“a car hitting the side of another car at an intersection”
Table 1: Prompt templates for collision type classification. Five prompts per class are averaged into one text embedding.
Figure 4: Distribution of ground-truth accident times (seconds) across the 2,211 synthetic videos. Most incidents occur within the first 10 seconds of the clip.
Figure 5: Collision type frequency in the synthetic development split. The rear-end category has 12 times more samples than the single-vehicle category.
Figure 6: Ground-truth impact point distribution across 2,211 synthetic videos, colored by collision type. Points cluster near the frame center, with head-on and sideswipe events showing the widest spatial spread.
Component
Parameter
Value
Temporal
Smoothing window w
5
Z-score threshold τ
1.5
Spatial
Start frame
centered on t∗
Context window T
30 frames
Pyramid scale
0.5
Pyramid levels
3
Table 2: Pipeline hyperparameters, held constant for all test videos.
Figure 7: Predicted collision type distribution on the 2,027 real test videos. Sideswipe and single dominate the predictions, while rear-end is rarely selected, inverting the synthetic training distribution.
We present SynCrash, a multi-stage pipeline for zero-shot accident detection, spatial localization, and collision-type classification in fixed-view CCTV surveillance video. Our approach addresses the ACCIDENT at CVPR 2026 Challenge, which requires predicting when an accident occurs, where in the frame the impact happens, and what type of collision it is, all without access to labeled real-world training data. The pipeline operates in three decoupled stages: (1) Temporal localization via a VideoMAEv2-giant backbone fine-tuned on CARLA-based synthetic clips with metadata-aware embeddings and dense sliding-window inference; (2) Spatial localization using YOLO for object detection combined with a physics-informed hybrid heuristic that leverages bounding-box overlap and trajectory-based reasoning to predict the impact point; and (3) Collision-type classification using a lightweight rule-based strategy derived from the number and configuration of detected vehicles. The key insight is that temporal understanding benefits from supervised fine-tuning on synthetic data, whereas spatial understanding is better served by pretrained object detectors and physics priors that transfer naturally across domains.
We present the 1st-place solution to the ACCIDENT challenge at the CVPR 2026 AUTOPILOT Workshop, which asks for zero-shot prediction of accident timing, impact centroid, and collision type from CCTV footage. On a frozen Qwen3-VL-32B-Instruct checkpoint we build a three-stage pipeline (full-video joint prediction, time refinement, and single-frame grounding of the impact centroid), run the same pipeline a second time on a 235B Mixture-of-Experts sibling, blend the two outputs 9:1, and finally snap each predicted point onto the nearest vehicle detection. The final system reaches Public LB 0.55469 / Private LB 0.57080, roughly +0.21 over the strongest host baseline (Molmo-7B, 0.358) and wins the challenge. We ablate each component, report the negative results that shaped the final design, and release the code at https://github.com/fuumin621/cvpr2026-accident-1st-place-solution.
Traffic surveillance cameras capture accidents continuously, yet converting raw CCTV footage into structured event records that pinpoint when, where, and what type of collision occurred remains unsolved at scale. The ACCIDENT @ CVPR benchmark evaluates exactly this joint prediction under a strict constraint: no labeled real-world training data is available. We introduce a training-free, two-pass coarse-to-fine pipeline that pairs a frozen Qwen3-VL-32B-Instruct vision-language model with YOLO11x object detection and BoT-SORT tracking. A first pass sparsely samples the full clip to anchor the collision moment in time; a second pass re-examines a tight window around that estimate using frames annotated with stable vehicle identities and normalized bounding-box coordinates, which gives the model both a visual overlay and an explicit numeric description of the same scene. On the official 2,027-clip real-CCTV test set, our system achieves a three-way harmonic mean score of 0.504, surpassing all organizer-published baselines including the best multi-model ensemble (0.412) by a 22% relative margin.
Dipit Saha, Shah Mohammad Abdul Mannan, Mohammad Raihan Rashid +2
Bangladesh University of Engineering and Technology