Source-linked AI summary
CADP: A Novel Dataset for CCTV Traffic Camera based Accident Analysis
Ankit Shah, Jean Baptiste Lamare, Tuan Nguyen Anh, Alexander Hauptmann
TL;DR
The paper addresses limited public data for spatio-temporal traffic-accident analysis from CCTV views. It introduces CADP, applies Faster R-CNN with contextual mining for small-object detection, and evaluates accident forecasting with an LSTM-based architecture. The reported results include improved detection and forecasting about two seconds before accidents at 80% recall.
Problem
Public datasets for learning traffic-camera accident detection and prediction are limited because accidents are rare and camera footage is legally difficult to obtain.
Method
The paper constructs the CADP dataset and combines Faster R-CNN with Context Mining and Augmented Context Mining, then feeds detector features into an accident-forecasting LSTM.
Results
+8.51% for CM and +6.20% for ACM are reported for object-detection accuracy, while the final forecasting model predicts accidents about 2 seconds before occurrence at 80% recall.
Takeaways & Limitations
CADP supplies CCTV traffic-accident data for spatio-temporal learning, and contextual features improve small-object detection while supporting accident forecasting.
Abstract
from arXiv · showhide
This paper presents a novel dataset for traffic accidents analysis. Our goal is to resolve the lack of public data for research about automatic spatio-temporal annotations for traffic safety in the roads. Through the analysis of the proposed dataset, we observed a significant degradation of object detection in pedestrian category in our dataset, due to the object sizes and complexity of the scenes. To this end, we propose to integrate contextual information into conventional Faster R-CNN using Context Mining (CM) and Augmented Context Mining (ACM) to complement the accuracy for small pedestrian detection. Our experiments indicate a considerable improvement in object detection accuracy: +8.51% for CM and +6.20% for ACM. Finally, we demonstrate the performance of accident forecasting in our dataset using Faster R-CNN and an Accident LSTM architecture. We achieved an average of 1.684 seconds in terms of Time-To-Accident measure with an Average Precision of 47.25%. Our Webpage for the paper is https://goo.gl/cqK2wE
1. Introduction
The paper introduces CADP, a publicly available CCTV traffic-camera dataset designed to support spatio-temporal accident analysis and forecasting. It also evaluates object detection and contextual enhancements for small objects.
- Traffic-camera views offer fixed, wider coverage and lower per-vehicle monitoring costs than first-person views.The paper connects these properties to road-safety monitoring with higher-quality cameras.
- Public datasets for traffic-camera accident learning are limited because accidents are rare and camera access is legally difficult.
- CADP contains 230 videos, 1,416 accident segments, and 205 HD segments with spatio-temporal annotations.The dataset supports temporal segmentation, object detection, tracking, collision detection, accident detection, and prediction.
- The paper applies Faster R-CNN and accident forecasting models, while testing Context Mining and Augmented Context Mining for small-object detection.
2. Related Work
Related work covers traffic-monitoring datasets, accident forecasting from dashboard cameras, and the evolution of deep-learning object detectors. The paper positions CCTV views as complementary because they overview complete roads and track more vehicles.
- Existing traffic-monitoring datasets include MIT, NGSIM, CBSR, CVRR, QMUL, and KIT resources with differing traffic-scene coverage and annotations.
- Dashboard Accident Dataset uses dashboard-camera videos for accident forecasting, while traffic cameras provide a broader road overview for tracking vehicles.
- Object-detection research progressed from proposal-based R-CNN methods to Fast R-CNN and end-to-end Faster R-CNN with CNN-based region proposals.
3. Car Accidents Dataset
CADP is a traffic-accident dataset with temporal and spatial annotations, many positive accident events, predominantly small objects, and varied real-world video conditions. Compared with related datasets, it emphasizes accident count and full spatio-temporal coverage.
- CADP is compared with related datasets using temporal annotation, spatial annotation, accident focus, CCTV capture, and positive-video counts.The comparison notes that CADP is not largest by hours but is largest by number of accidents.
- A major portion of CADP consists of small objects, creating a challenging surveillance-video detection setting.
- CADP contains 1,416 positive accident videos, compared with 151 road-accident videos in UCF-Crimes and about 600 in DAD.
- The dataset includes YouTube videos captured with varied camera types and qualities, weather conditions, and editing or resampling.
4. Improved Faster R-CNN and Forecasting
The paper improves Faster R-CNN for small-object detection by mining contextual regions, then uses detector features with an attention-based LSTM for accident forecasting.
- Improved Faster R-CNN for Object Detection: Faster R-CNN uses a two-stage pipeline that generates region proposals internally before classifying and regressing proposed regions.The Region Proposal Network reduces dependence on external proposal algorithms.
- Context Mining: Small pedestrian regions in CADP suffer degraded detection accuracy, motivating context mining around object proposals.The dataset contains many objects smaller than 100 pixels, and pedestrians occupy fewer pixels than vehicles.
- Context Mining: Context Mining extends small-object regions and selects informative pooled contextual features with a Maxout network.The mined context region contains the original object region, while Maxout selects the strongest responses from pooled contextual features.
- Context Mining: Augmented Context Mining extends or narrows small bounding boxes horizontally and vertically to expose diverse contextual patterns.This approach addresses inaccurate annotations and possible inclusion of other objects when boxes are enlarged.
- Context Mining: Context mining is applied only when the bounding-box area is at most 1% of the image area.The method uses α = 0.01 as the area-ratio threshold.
- Accident Forecasting: The accident-forecasting system feeds 2048-dimensional Faster R-CNN features into Dynamic-Spatial-Attention LSTM to produce accident scores over time.The attention mechanism distributes weights across detected spatial objects and processes frame sequences dynamically.
5. Results
Experiments compare object detectors, evaluate Faster R-CNN across folds, test contextual extensions for pedestrian detection, and assess accident forecasting. Faster R-CNN is selected as baseline, while CM improves detection and forecasting results.
- Object detection: Faster R-CNN performed stably across vehicle categories but degraded and became unstable for the Person category.The cross-validation measure was mAP@0.5.
- Object detection: Faster R-CNN was selected as the baseline after SSD showed approximately 19.69% lower mAP@0.5.The comparison motivated further experimentation with Faster R-CNN.
- Context Mining: +8.51% for CM and +6.20% for ACM over the baseline were reported with appropriate contextual-mining settings.The selected CM settings were nc = 16 and s = 4.
- Context Mining: CM outperformed ACM by about two points in mAP@0.5 for pedestrian detection.The results indicate that mining fewer contexts and gradually extending regions performed better.
- Accident forecasting: 1.684 seconds was the reported accident-forecasting result for Context Mining, while the corresponding 80%-recall ToA was 3.078 seconds.Using CM features led to better forecasting results.
6. Conclusion
The conclusion presents CADP as a CCTV traffic-camera dataset and reports object-detection and accident-forecasting evaluations. Context mining improves the initial results, while gradual context mining outperforms augmented context mining for object detection.
- Dataset and evaluation: CADP is introduced as a dataset from CCTV traffic-camera videos for traffic accident analysis.The dataset supports evaluation of object detection and accident forecasting models.
- Object detection: Context mining and augmented context mining improve the initial object-detection results on CADP.The conclusion describes these methods as additions to the baseline models.
- Object detection: Gradual context mining achieves a better object-detection score than augmented context mining.The conclusion states that augmented context mining does not improve the score obtained with gradual context mining.
- Accident forecasting: The final accident-forecasting model predicts accidents about 2 seconds before they occur with 80% recall.This is the reported forecasting outcome in the conclusion.
A.1. Data collection
The data-collection process addresses the difficulty of obtaining public traffic-accident data from fixed third-person camera views. The main constraints are accident rarity and limited access to traffic-camera footage.
- Collection challenges: Traffic accidents are rare, making it infeasible to wait for an accident at a live road-intersection camera.The rarity constraint is identified as an abnormality challenge.
- Collection challenges: Access to traffic-camera data is often limited, restricting public availability of fixed third-person accident views.The passage identifies access as the second major collection challenge.
A.2. Annotations
CADP annotations were created through a two-stage process that first isolates relevant CCTV segments and then applies spatio-temporal annotation. This addresses the fact that accident videos often contain only brief relevant segments.
- Temporal segmentation: The annotation process first extracted starting and ending timestamps for CCTV traffic-camera segments from each video.Human annotators performed the temporal-segmentation stage.
- Spatio-temporal annotation: The extracted segments were then collected and spatio-temporally annotated using the VATIC tool.This formed the second stage of the annotation process.
B. Quality assurance
The dataset annotation process used annotators, expert reviewers, guidelines, pretraining, and staged review to improve quality and confidence. Initial trial performance was poor, but quality improved after feedback and standardized workflows.
- Annotation setup: Two human annotators performed spatio-temporal annotation with VATIC, while additional quality practices were applied after low trial-work quality was observed.The quality-control process was introduced to improve confidence in submitted annotations.
- Peer-review: Three expert reviewers screened submitted annotations, rejecting clearly poor work, accepting clearly good work, and investigating borderline cases.Reviewers could also modify borderline annotations according to the guideline.
- Annotation guideline: An 8-page guideline aligned reviewers and annotators on object definitions, procedures, and examples of bad work.The guideline was created to address misunderstandings observed during the trial period.
- Documentation: The annotation guideline and a separate process report were made publicly available, alongside the paper’s reported quality-assurance insights.These materials document the annotation process beyond the summarized quality-control steps.
- Pretraining: Trial-time performance was initially far below standard, with a reviewer reporting that barely 1% of submitted work could be accepted.The trial served as pretraining for annotators unfamiliar with spatio-temporal annotation.
- Process outcomes: After performance reached the standard, annotation quality improved and annotators began reporting software and server problems to reviewers.This feedback shifted part of the collaboration toward identifying technical issues.
- Workflow control: To prevent concurrent edits from overwriting unsaved work, annotators submitted completed work before reviewers received non-overlapping assigned episodes.Reviewers returned poor annotations for modification with error lists.
- Annotation staffing: Although stronger paid solutions such as AMT were available, the team used voluntary or nearly unpaid annotation because the workload was limited and cost savings were desired.Two annotators and three expert reviewers participated voluntarily under the described arrangement.