Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

AccidentBlip: Agent of Accident Warning based on MA-former

Yihua Shao, Yeling Xu, Xinwei Long, Siyu Chen, Ziyang Yan, Yang Yang, Haoting Liu, Yan Wang, Hao Tang, Zhen Lei

TL;DR

AccidentBlip introduces a vision-only framework for accident detection and prediction in complex traffic by leveraging a Motion Accident Transformer (MA-former) that replaces Q-former self-attention with temporal attention. The method processes temporal multi-view video frames, fusing information across cameras and vehicles through residual connections and multi-query aggregation to produce frame-level danger assessments. On the DeepAccident dataset, AccidentBlip achieves state-of-the-art precision and accident prediction accuracy in both ego-centric and V2X scenarios, with effective early warnings in horizons up to $2$–$3$ seconds (decreasing beyond $3$–$4$ seconds). The approach avoids heavy language models and expensive multimodal fusion, offering a practical, vision-only solution with strong potential for real-world autonomous driving safety systems.

Abstract

In complex transportation systems, accurately sensing the surrounding environment and predicting the risk of potential accidents is crucial. Most existing accident prediction methods are based on temporal neural networks, such as RNN and LSTM. Recent multimodal fusion approaches improve vehicle localization through 3D target detection and assess potential risks by calculating inter-vehicle distances. However, these temporal networks and multimodal fusion methods suffer from limited detection robustness and high economic costs. To address these challenges, we propose AccidentBlip, a vision-only framework that employs our self-designed Motion Accident Transformer (MA-former) to process each frame of video. Unlike conventional self-attention mechanisms, MA-former replaces Q-former's self-attention with temporal attention, allowing the query corresponding to the previous frame to generate the query input for the next frame. Additionally, we introduce a residual module connection between queries of consecutive frames to enhance the model's temporal processing capabilities. For complex V2V and V2X scenarios, AccidentBlip adapts by concatenating queries from multiple cameras, effectively capturing spatial and temporal relationships. In particular, AccidentBlip achieves SOTA performance in both accident detection and prediction tasks on the DeepAccident dataset. It also outperforms current SOTA methods in V2V and V2X scenarios, demonstrating a superior capability to understand complex real-world environments.

AccidentBlip: Agent of Accident Warning based on MA-former

TL;DR

AccidentBlip introduces a vision-only framework for accident detection and prediction in complex traffic by leveraging a Motion Accident Transformer (MA-former) that replaces Q-former self-attention with temporal attention. The method processes temporal multi-view video frames, fusing information across cameras and vehicles through residual connections and multi-query aggregation to produce frame-level danger assessments. On the DeepAccident dataset, AccidentBlip achieves state-of-the-art precision and accident prediction accuracy in both ego-centric and V2X scenarios, with effective early warnings in horizons up to – seconds (decreasing beyond – seconds). The approach avoids heavy language models and expensive multimodal fusion, offering a practical, vision-only solution with strong potential for real-world autonomous driving safety systems.

Abstract

In complex transportation systems, accurately sensing the surrounding environment and predicting the risk of potential accidents is crucial. Most existing accident prediction methods are based on temporal neural networks, such as RNN and LSTM. Recent multimodal fusion approaches improve vehicle localization through 3D target detection and assess potential risks by calculating inter-vehicle distances. However, these temporal networks and multimodal fusion methods suffer from limited detection robustness and high economic costs. To address these challenges, we propose AccidentBlip, a vision-only framework that employs our self-designed Motion Accident Transformer (MA-former) to process each frame of video. Unlike conventional self-attention mechanisms, MA-former replaces Q-former's self-attention with temporal attention, allowing the query corresponding to the previous frame to generate the query input for the next frame. Additionally, we introduce a residual module connection between queries of consecutive frames to enhance the model's temporal processing capabilities. For complex V2V and V2X scenarios, AccidentBlip adapts by concatenating queries from multiple cameras, effectively capturing spatial and temporal relationships. In particular, AccidentBlip achieves SOTA performance in both accident detection and prediction tasks on the DeepAccident dataset. It also outperforms current SOTA methods in V2V and V2X scenarios, demonstrating a superior capability to understand complex real-world environments.
Paper Structure (13 sections, 14 equations, 4 figures, 4 tables)

This paper contains 13 sections, 14 equations, 4 figures, 4 tables.

Table of Contents

  1. INTRODUCTION
  2. RELATED WORK
  3. End-to-End Automatic Driving
  4. Accident Detection and Prediction
  5. METHODOLOGY
  6. Overall Architecture
  7. The Proposed MA-former
  8. Complex Systems Adaptation
  9. EXPERIMENTS
  10. Experiment Setting
  11. Main Results
  12. Ablation Study
  13. CONCLUSIONS

Figures (4)

  • Figure 1: Overall architecture of AccidentBlip. We input the multi-view images of each frame into ViT to obtain the features of each frame. Then we input previous frame query and multi-view images features into MA-former and input the text description. Finally, we use residual to connect the previous frame query and the image features of this frame with MA-former's output and input it into the MA-former corresponding to the next frame.
  • Figure 2: Architecture of MA-former. MA-former replaces Q-former's self-attention with temporal attention to realize processing temporal images.
  • Figure 3: Architecture of multi-vehicle sensing system. We realize the fusion of multi-vehicle information by merging the queries output from each on-vehicle MA-former.
  • Figure 4: Example of interaction between AccidentBlip and driver. AccidentBlip assesses the risk of accidents in the surrounding environment and alerts the driver to drive with caution.