ACCIDENT：交通监控视频中事故检测、定位与分类的模块化零样本流水线

一、前言

ACCIDENT不但是个开源的数据集，它还提供了baseline
项目地址：https://accidentbench.github.io/
代码地址：https://github.com/accidentbench/ACCIDENT

二、ACCIDENT：交通监控视频中事故检测、定位与分类的模块化零样本流水线

论文基本信息

论文核心内容

该论文提出了一种模块化零样本（Zero-Shot）流水线，用于在交通监控视频中自动检测交通事故，无需任何标注的真实世界训练数据。该方法面向 ACCIDENT @CVPR 2026 竞赛，要求对视频预测三个目标：

1. 事故时间（Temporal Localization）：通过计算相邻帧的亮度差异，构建一维信号，利用 z-score 归一化和峰值检测定位碰撞发生的时刻。
2. 碰撞位置（Spatial Localization）：在检测到的时间点附近，使用 Farneback 稠密光流算法累积运动幅度图，通过 90% 分位数阈值和加权质心计算碰撞点的归一化坐标。
3. 碰撞类型分类（Collision Type Classification）：利用预训练的 CLIP（ViT-B/32）模型，将事故帧与五类自然语言描述（head-on、rear-end、sideswipe、single-vehicle、t-bone）进行余弦相似度匹配，实现零样本分类。

数据集与结果

• 开发集：2,211 个由 CARLA 模拟器生成的合成视频
• 测试集：2,027 个真实监控摄像头录像
• 公开榜分数：在真实 CCTV 测试集上取得 0.2523（基于约 25% 测试数据）

整个流水线无需任何领域特定的微调，仅使用预训练模型权重，在 Kaggle 平台的 NVIDIA T4 GPU 上运行，处理全部测试视频约需 2 小时。

摘要

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.
我们描述了一个为ACCIDENT @ CVPR 2026挑战赛开发的零样本（zero-shot）流水线。该挑战赛要求在监控视频中预测交通事故发生的时间、地点和类型，且不使用标注的真实世界训练数据。我们的方法将问题分解为三个独立的模块。

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.
第一个模块通过对z分数归一化的帧差信号运行峰值检测来在时间上定位碰撞。第二个模块通过使用Farneback算法计算累积密集光流幅度图的加权质心来找到撞击位置。

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 finetuning is involved; the pipeline processes each video using only pre-trained model weights.
第三个模块通过测量检测到的峰值附近帧的CLIP图像嵌入与从每个碰撞类别的多提示自然语言描述构建的文本嵌入之间的余弦相似度来分类碰撞类型。没有涉及特定领域的微调；该管道仅使用预训练的模型权重处理每个视频。

Our implementation is publicly available as a Kaggle notebook [Thakur and Talele, 2026].
我们的实现可在 Kaggle Notebook 中公开获取 [Thakur and Talele, 2026]。

1 引言

Road traffic crashes kill over one million people each year according to the World Health Organization [World Health Organization, 2023]. Surveillance cameras installed at intersections and along highways record continuous footage that captures the circumstances of many of these incidents. If this footage could be analyzed automatically, emergency dispatchers would receive faster alerts and investigators would gain access to objective scene reconstructions.
根据世界卫生组织的数据，每年有超过一百万人死于道路交通事故[World Health Organization, 2023]。安装在交叉路口和沿高速公路的监控摄像头会记录连续的视频片段，捕捉到许多事故发生时的情景。如果能够自动分析这些视频片段，紧急调度员将能更快地收到警报，调查人员也能获得客观的现场重建信息。

Yet most current detection methods require supervised training on annotated video from the same deployment environment [Sultani et al., 2018, Yao et al., 2022], which makes them expensive to transfer across camera installations with different viewpoints, lighting, and traffic patterns.
然而，目前大多数检测方法都需要在相同部署环境的标注视频上进行监督训练[Sultani et al., 2018, Yao et al., 2022]，这使得它们在跨越不同视角、光照和交通模式的摄像头安装时转移成本高昂。

The ACCIDENT @ CVPR 2026 competition [Picek et al., 2026] tests whether accident analysis can work without any labeled real-world data. Participants receive only synthetic videos generated by the CARLA simulator [Dosovitskiy et al., 2017] for development (see Figure 1); the test set is composed of real CCTV recordings and manual annotation of it is prohibited.
ACCIDENT @ CVPR 2026 竞赛 [Picek 等人, 2026] 测试了在没有任何标记的真实世界数据的情况下，事故分析是否可行。参赛者仅接收由 CARLA 模拟器 [Dosovitskiy 等人, 2017] 生成的合成视频用于开发（参见图 1)；测试集由真实的闭路电视录像组成，禁止对其进行手动标注。

Three predictions are required per video: (1) the accident time in seconds, (2) normalized spatial coordinates of the impact point, and (3) the collision type from a five-class taxonomy (head-on, rear-end, sideswipe, single-vehicle, t-bone). Scoring uses the harmonic mean of a Gaussian temporal similarity, a Gaussian spatial similarity, and top-1 classification accuracy, so a poor prediction in any one dimension pulls the overall score down.
每个视频需要三个预测：(1) 事故发生时间（秒），(2) 撞击点的归一化空间坐标，以及 (3) 来自五类分类的碰撞类型（正面碰撞、追尾碰撞、侧面碰撞、单车事故、T字形碰撞）。评分采用高斯时间相似度、高斯空间相似度以及 Top-1 分类准确率的调和平均值，因此任何一个维度的预测不佳都会拉低整体分数。

We address each target with a separate module. Temporal localization runs statistical anomaly detection on pixelintensity differences between consecutive frames. Spatial localization accumulates dense optical flow magnitudes and extracts a weighted centroid. Collision classification passes frames near the detected peak through CLIP [Radford et al., 2021], a contrastive vision-language model trained on 400 million image-text pairs, and selects the class whose text prompts produce the highest cosine similarity. Because the modules are independent, any one of them can be swapped or tuned without modifying the rest.
我们为每个目标配备一个独立的模块。时间定位在连续帧之间的像素强度差异上运行统计异常检测。空间定位累积密集光流幅度并提取加权质心。碰撞分类将检测到的峰值附近的帧通过CLIP [Radford et al., 2021] 进行处理，CLIP是一个在4亿个图像-文本对上训练的对比视觉-语言模型，并选择其文本提示产生最高余弦相似度的类别。由于模块是独立的，因此可以替换或调整其中任何一个模块，而无需修改其他模块。

The full pipeline is implemented in a public Kaggle notebook [Thakur and Talele, 2026].
整个流程在公开的Kaggle笔记本[Thakur and Talele, 2026]中实现。

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.
图 1：来自 ACCIDENT @ CVPR 2026 数据集中合成的 CARLA 交通事件的时间顺序采样帧。大多数剪辑以 20 FPS 的分辨率记录约 18 秒的运动。