Enhancing autonomous vehicle hyperawareness in busy traffic environments: A machine learning approach

Area of Science:

• Artificial Intelligence
• Transportation Engineering
• Robotics

Background:

• Autonomous vehicles (AVs) are increasingly integrated into urban environments, necessitating advanced safety protocols.
• Real-time identification of traffic conflicts is crucial for AV operational safety and the protection of road users.
• Existing algorithms require continuous improvement to handle complex urban traffic dynamics.

Purpose of the Study:

• To develop and evaluate a machine learning model for predicting traffic conflicts involving AVs and surrounding road users.
• To utilize AV sensor data for real-time hazard detection and categorization.
• To assess the transferability of the trained model to new datasets and locations.

Main Methods:

• Collected AV data from multiple US cities.
• Developed a machine learning model using Isolation Forest - Convolutional Neural Network - Long Short-Term Memory (IF-CNN-LSTM) layers.
• Trained the model on real-time road user trajectories and headings to predict conflicts three seconds ahead.
• Evaluated model performance and transferability across different datasets.

Main Results:

• The IF-CNN-LSTM model achieved high accuracy in predicting conflict occurrence and severity (sensitivity=83.5%, fallout=11%).
• AV conflict prediction sensitivity ranged from 89% to 95%, outperforming existing models.
• Real-time prediction of hazardous conflicts for surrounding road users showed high sensitivity (82%-87%).
• The model retained nearly 97% of its performance when transferred with sufficient tuning data.

Conclusions:

• The proposed model demonstrates significant potential for enhancing AV safety through accurate, real-time conflict prediction.
• Onboard AV sensors can be effectively utilized for real-time safety applications.
• The model's transferability suggests adaptability to diverse urban environments with minimal retraining effort.