Prediction of jaywalker-vehicle conflicts based on encoder-decoder framework utilizing multi-source data.

Area of Science:

• Road safety engineering
• Artificial intelligence in transportation
• Traffic conflict analysis

Background:

• Jaywalker-vehicle (J-V) conflicts are frequent and hazardous at mid-block locations without designated crossings in China.
• The unpredictable nature of jaywalking necessitates predictive models for timely warnings to road users.

Purpose of the Study:

• To introduce a novel encoder-decoder framework for predicting J-V conflict severity using multi-source data.
• To enhance road safety by providing pre-conflict warnings through accurate severity prediction.

Main Methods:

• A novel encoder-decoder framework was developed, incorporating three encoders: J-V interaction (Bi-LSTM), jaywalker motion (Bi-LSTM), and background information (MLP).
• Features from these encoders were concatenated and fed into a conflict severity decoder (MLP) for prediction.
• A case study was conducted using video data from three mid-block locations in Nanjing, China.

Main Results:

• The proposed encoder-decoder model outperformed classical models in predictive metrics.
• Incorporating background information significantly improved the model's performance, with an average enhancement of 24.291% across four evaluative metrics.
• Transferability analysis showed that predictive metrics stabilized between 80% and 95% when 40%-50% of data from a new location was included.

Conclusions:

• The developed encoder-decoder model offers a robust solution for predicting J-V conflict severity.
• Background information is a critical factor in improving the accuracy of J-V conflict prediction models.
• The model demonstrates good transferability, suggesting its potential for widespread application in enhancing jaywalking safety.