Source-linked AI summary
SynCrash: A Multi-Stage Pipeline for Zero-Shot Accident Detection and Localization in Traffic Surveillance Video
Arkya Jyoti Bagchi, Ritul Jangir, Varun Raskar
TL;DR
SynCrash targets zero-shot accident understanding in fixed-view CCTV, where systems must predict when an accident occurs, where impact happens, and what collision type occurred without labeled real training data. Its modular pipeline combines synthetic-data temporal learning, detector- and physics-based spatial reasoning, and rule-based classification; it ranks 17th on the private leaderboard. The main boundary is that detection quality, fixed-length clips, simple classification heuristics, and absent stage feedback constrain robustness and joint optimization.
Problem
The ACCIDENT challenge requires joint prediction of accident time, impact location, and collision type without labeled real-world training data.
Method
SynCrash decomposes the task into temporal localization with VideoMAEv2, physics-informed spatial localization with YOLO, and rule-based collision classification.
Results
SynCrash ranks 17th overall on the private leaderboard and provides a strong, efficient baseline under the training-free setting.
Takeaways & Limitations
The experiments support modular decomposition because temporal, spatial, and classification components exhibit different sensitivities to domain shift.
Takeaways & Limitations
The method remains vulnerable to noisy detections, very short or gradual events, fine-grained multi-vehicle interactions, and extreme environmental conditions.
Abstract
from arXiv · showhide
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.
1. Introduction
SynCrash addresses zero-shot accident understanding in difficult fixed-view CCTV by predicting accident time, impact location, and collision type under synthetic-to-real constraints. It decomposes the task into modular stages and assigns each subproblem an appropriate tool.
- Fixed-view CCTV accident analysis is difficult because footage often has low resolution, compression artifacts, occlusions, wide fields of view, and variable lighting.
- The ACCIDENT challenge requires predicting accident time, normalized impact coordinates, and collision type from CCTV clips.
- The benchmark evaluates all three outputs jointly through a harmonic mean, while real labeled training data are unavailable.
- SynCrash decomposes accident understanding into temporal, spatial, and classification stages, each addressed with a task-specific tool.
2. Method
SynCrash uses a modular pipeline that first localizes accident time, then estimates the impact point, and finally classifies collision type. The design combines synthetic-data temporal learning with object-centric and physics-informed spatial reasoning.
- SynCrash separates temporal localization, spatial localization, and collision-type classification instead of learning the outputs jointly.
- Temporal Localization: Temporal localization uses VideoMAEv2 on short CCTV clips, augmented with scene, weather, and time-of-day metadata embeddings.
- Temporal Localization: The model is trained on synthetic CARLA clips and uses CCTV-style degradation augmentations before dense sliding-window inference.
- Spatial Localization: Spatial localization detects vehicles with YOLO and estimates short-term object motion from matched detections.
- Spatial Localization: A priority-based heuristic uses bounding-box overlap, trajectory intersections, weighted object centers, and fallback midpoints to infer the collision point.
- Collision Classification: Collision type is assigned by rules based on detected-vehicle count, spatial configuration, and relative motion.
3. Experimental Setup
The experiments follow the ACCIDENT challenge’s zero-shot sim-to-real protocol, training on synthetic CARLA data and evaluating on real CCTV footage. Performance is summarized by temporal, spatial, and collision-type components combined through a harmonic mean.
- Models are trained only on synthetic CARLA traffic videos and evaluated on real CCTV footage without labeled real training samples.
- The official metric combines temporal accuracy, spatial localization accuracy, and collision-type classification accuracy.
- The final benchmark score is the harmonic mean of the three component scores.
- Temporal and spatial scores use Gaussian-based similarity functions over prediction errors, while classification uses top-1 accuracy.
4. Results and Analysis
SynCrash provides a competitive and efficient modular baseline, ranking 17th on the private leaderboard. Design analyses favor decoupled temporal and spatial reasoning with object-centric and physics-based representations over joint or more complex alternatives.
- Main Results: SynCrash achieves competitive overall performance on both public and private ACCIDENT leaderboards under the official harmonic-mean metric.
- Main Results: The approach ranks 17th overall on the private leaderboard while serving as a strong and efficient training-free baseline.
- Effect of Design Choices: A joint ViViT multi-task model showed poor performance, unstable spatial predictions, and classification bias, indicating weak task disentanglement.
- Effect of Design Choices: VideoMAEv2 improved temporal localization, while YOLO with physics-informed reasoning produced more accurate and stable spatial localization than Grad-CAM, especially for multi-vehicle scenes.
- Effect of Design Choices: RAFT and graph-based interaction models offered useful insights but did not consistently outperform the simpler modular design.
- Effect of Design Choices: The experiments support decoupling temporal and spatial reasoning and using object-centric representations with physics-based heuristics under zero-shot sim-to-real constraints.
5. Discussion and Lessons Learned
SynCrash’s modular design matches different task sensitivities under zero-shot sim-to-real conditions: learned temporal representations transfer useful motion cues, while object-centric and physics-informed reasoning supports spatial localization. The experiments also indicate that decoupling specialized components is more robust than tightly coupled or complex alternatives in degraded CCTV video.
- Temporal localization: Temporal localization benefits from learned representations because approaching vehicles and abrupt deceleration patterns transfer reasonably well from synthetic to real data.VideoMAEv2 captures these spatiotemporal cues and enables reliable accident-timing prediction.
- Spatial localization: Spatial localization is highly sensitive to domain gap, so YOLO and physics-informed reasoning exploit geometric relationships instead of fragile visual mappings.The approach uses object-centric representations and domain-invariant geometric relationships under differences in camera geometry, resolution, and compression.
- Modular design: Specialized components operate in their respective strengths: VideoMAEv2 supplies temporal understanding, while object detectors provide more accurate localization.Decoupling avoids interference from competing objectives because the backbone has coarse spatial signals and detectors lack temporal context.
- Physics-informed reasoning: Overlap, trajectory intersection, and approach-velocity heuristics are more stable under noisy, low-resolution CCTV conditions because they rely on geometric consistency.These physics-based cues use less appearance-dependent information than learned visual features.
- Overall lesson: Under training-free constraints, combining learned temporal representations with object-centric detection and physics-informed reasoning offers a more reliable alternative to fully end-to-end approaches.The conclusion follows from the distinct domain sensitivities of temporal and spatial prediction.
- Alternative approaches: Joint multi-task learning struggled to balance competing objectives, producing unstable spatial predictions and classification bias toward dominant classes.The finding supports separating heterogeneous tasks under domain shift rather than tightly coupling them in a shared backbone.
- Alternative approaches: Graph-based interaction models depended heavily on accurate tracking and stable detections, which are difficult to obtain in low-quality surveillance videos.These requirements limit the reliability of complex relational modeling in degraded visual conditions.
- Practical lessons: Overall, methods relying heavily on dense motion estimation or complex relational modeling were less robust in degraded visual conditions than simpler modular approaches using domain-invariant reasoning.The experiments reinforce modularity and geometric reasoning as practical choices for zero-shot sim-to-real accident understanding.
6. Limitations
SynCrash remains constrained by heuristic classification, detection-dependent spatial localization, fixed-length temporal clips, sequential stages, and limited robustness under extreme domain conditions.
- Classification: Collision-type classification struggles with fine-grained multi-vehicle interactions, especially under noisy detections.The classifier relies on simple heuristics rather than richer interaction modeling.
- Spatial localization: Spatial localization depends on detection quality and is vulnerable in crowded or low-resolution scenes.This creates a practical boundary for impact-point prediction in visually degraded settings.
- Temporal localization: Fixed-length clips reduce temporal sensitivity to very short or gradual accident events.The limitation affects the temporal-localization stage rather than the spatial or classification stages.
- Pipeline integration: The sequential pipeline provides no feedback between stages, limiting joint optimization.The stages are therefore not jointly refined using downstream information.
- Domain scope: Although robust to sim-to-real conditions, the method lacks explicit domain adaptation and may degrade under extreme lighting, weather, or viewpoints.The authors identify interaction modeling, detection robustness, and tighter stage integration as future directions.