Source-linked AI summary
Uncertainty-based Traffic Accident Anticipation with Spatio-Temporal Relational Learning
Wentao Bao, Qi Yu, Yu Kong
TL;DR
Existing accident-anticipation methods overlook spatio-temporal relations and predictive uncertainty, despite the safety importance of early, reliable predictions. The paper combines relational learning with Bayesian uncertainty modeling and reports superior performance across evaluated datasets, including earlier anticipation on DAD.
Problem
Existing methods typically ignore relations between accident-relevant agents and uncertainty estimation, both important for safety-guaranteed systems.
Method
The model learns spatio-temporal relational features using graph convolution and recurrent networks, while Bayesian neural networks model uncertainty in latent relational representations.
Results
The model outperforms existing methods on DAD, A3D, and CCD; on DAD, it anticipates accidents 3.53 seconds earlier on average and reaches 72.22% average precision at the best precision setting.
Takeaways & Limitations
The uncertainty-based ranking loss improves relational-feature quality and performance, while Bayesian modeling also provides interpretable predictive uncertainty; CCD adds environmental and accident-reason annotations.
Takeaways & Limitations
High TTA can result from false positives, so the evaluation mainly reports TTA when the highest AP is achieved rather than relying on recall alone.
Abstract
from arXiv · showhide
Traffic accident anticipation aims to predict accidents from dashcam videos as early as possible, which is critical to safety-guaranteed self-driving systems. With cluttered traffic scenes and limited visual cues, it is of great challenge to predict how long there will be an accident from early observed frames. Most existing approaches are developed to learn features of accident-relevant agents for accident anticipation, while ignoring the features of their spatial and temporal relations. Besides, current deterministic deep neural networks could be overconfident in false predictions, leading to high risk of traffic accidents caused by self-driving systems. In this paper, we propose an uncertainty-based accident anticipation model with spatio-temporal relational learning. It sequentially predicts the probability of traffic accident occurrence with dashcam videos. Specifically, we propose to take advantage of graph convolution and recurrent networks for relational feature learning, and leverage Bayesian neural networks to address the intrinsic variability of latent relational representations. The derived uncertainty-based ranking loss is found to significantly boost model performance by improving the quality of relational features. In addition, we collect a new Car Crash Dataset (CCD) for traffic accident anticipation which contains environmental attributes and accident reasons annotations. Experimental results on both public and the newly-compiled datasets show state-of-the-art performance of our model. Our code and CCD dataset are available at https://github.com/Cogito2012/UString.
1 INTRODUCTION
Traffic accident anticipation is important for safety-guaranteed autonomous driving but remains difficult because cluttered scenes and limited cues obscure accident-relevant information. The proposed model learns agent relations and predictive uncertainty, and is evaluated on public datasets and a new annotated Car Crash Dataset.
- Accident anticipation predicts crashes from dashcam video before they happen, potentially giving self-driving systems time for urgent safety control.
- Cluttered traffic scenes and limited visual cues can overwhelm accident-relevant information with irrelevant objects.
- GCNs learn spatial relations while RNNs learn temporal relations, with Bayesian neural networks addressing predictive uncertainty.
- The model captures relational features and predictive uncertainty beyond agent-specific features used by existing RNN-based methods.
- 3.53 seconds earlier anticipation and 72.22% average precision are reported on DAD, while CCD adds environmental and accident-reason annotations.
- The model jointly uses agent-specific features, spatio-temporal relations, and predictive uncertainty for accident anticipation.
2 RELATED WORK
Prior accident-anticipation work commonly overlooks relations among accident-relevant agents and does not estimate uncertainty. Related uncertainty methods model latent representations or network weights probabilistically, providing context for this paper's relational and Bayesian design.
- Existing accident-anticipation methods typically ignore relations between accident-relevant agents despite their importance for anticipating future accidents.
- Existing methods also omit uncertainty estimation, which the paper identifies as critical for safety-guaranteed systems.
- The proposed framework represents traffic scenes with graph-embedded object features and learns latent relational representations through coupled GCN and RNN processing.
- Prior uncertainty approaches model latent relational observations or neural-network weights as random variables, including VAE-, VRNN-, and BNN-based formulations.
3 PROPOSED METHOD
The proposed method models accident-relevant agents and their coupled spatial-temporal relations from dashcam videos, while Bayesian prediction estimates uncertainty in accident scores. It combines graph convolution, recurrent modeling, and self-attention-based aggregation to improve relational representation learning.
- Spatio-Temporal Relational Learning: The framework fuses relational features with agent-specific features and feeds them into an RNN to update the hidden state over time.This couples spatial and temporal relation learning in a cyclic process while retaining both feature types.
- Spatio-Temporal Relational Learning: The model represents detected traffic objects as graph nodes and uses complete-graph adjacency matrices weighted by object-region distances.Closer object regions receive larger graph-convolution weights, although pixel distance may not reflect physical distance under occlusion.
- Spatio-Temporal Relational Learning: Graph convolution learns spatial relations from object features and recurrent hidden states, producing relational features that incorporate temporal context.The paper uses two stacked GCN layers, with the RNN hidden state fused into spatial relational learning.
- BNNs for Accident Anticipation: Bayesian neural networks replace deterministic score prediction so sampled network parameters can represent predictive uncertainty from latent relational features.The BNN module receives Z_t and uses variational inference with Bayes-by-Backprop to approximate the posterior over parameters.
- Temporal Self-Attention Aggregation: Self-attention aggregation adaptively combines hidden states from all time steps to predict a video-level accident score during training.The SAA layer is an auxiliary training objective and is not used during testing.
4 EXPERIMENTAL RESULTS
Experiments evaluate the model on CCD, DAD, and A3D using anticipation, uncertainty, ablation, and model-size analyses. The model anticipates accidents earlier on DAD, while uncertainty-based ranking and relational components contribute to performance.
- Datasets and evaluation: Experiments use the collected CCD dataset and the public DAD and A3D datasets, with state-of-the-art comparisons and ablation studies.CCD includes environmental and accident-reason annotations, while DAD and A3D provide public benchmarks.
- Datasets and evaluation: Average Precision measures accident-identification correctness, while mTTA and TTA@0.8 measure anticipation earliness at different recall conditions.The evaluation also reports mean aleatoric and epistemic uncertainty.
- Performance evaluation: 3.53 seconds earlier anticipation is achieved on DAD, with 53.7% AP compared with L-RAI and adaLEA.The model achieves the best mTTA on DAD while maintaining competitive AP.
- Performance evaluation: For recall rates larger than 80%, the method performs poorly compared with DSA, while high recall can also produce false alarms that reduce AP.This motivates evaluating models at their best AP rather than maximizing TTA alone.
- Ablation study: The uncertainty-based ranking loss contributes about 7.6% performance gain, while feature fusion between relational and agent-specific features contributes approximately 7%.Ablations also validate self-attention aggregation and the superiority of GCN over naive fully connected layers.
- Model size: The proposed model is much lighter than DSA and only slightly increases model size relative to its other variants.Parameter counts are reported in millions.
5 CONCLUSION
The paper presents uncertainty-based traffic accident anticipation with spatio-temporal relational learning and releases the CCD dataset. Its Bayesian formulation supports both improved anticipation performance and interpretation of predictive uncertainty.
- Conclusion: The model combines spatio-temporal relational learning with Bayesian uncertainty estimation to handle relational features and anticipation uncertainty from video data.The conclusion identifies uncertainty-based ranking loss as a source of performance improvement and interpretable uncertainty.
- Conclusion: The released CCD dataset contains rich environmental attributes and accident reason annotations for traffic accident anticipation.These annotations extend the dataset resources available for this task.