Multi-Object Trajectory Prediction Based on Lane Information and Generative Adversarial Network.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Current trajectory prediction algorithms struggle with real-world traffic complexity, leading to significant prediction errors.
• Existing methods often lack robustness in challenging environments like crowded scenes or poor lighting.

Purpose of the Study:

• To develop a robust multi-object trajectory prediction algorithm that overcomes limitations of current methods.
• To improve both lane detection accuracy and trajectory prediction precision using novel feature extraction and fusion techniques.

Main Methods:

• A Channel Attention-based Hybrid Dilated Convolution (CA-HDC) module for enhanced lane feature extraction.
• Integration of a lane information fusion module and a foresight-based trajectory adjustment module.
• Utilizing Socially acceptable trajectory with Generative Adversarial Networks (S-GAN) to minimize prediction errors.

Main Results:

• Improved lane detection accuracy in complex scenarios (crowded, shadow, arrow, crossroad, night) on the CULane dataset, with a 4.1% increase in average F1-measure compared to PINet.
• Reduced average displacement error by 4.27% and final displacement error by 7.53% in trajectory prediction tests on the D²-City dataset.
• Demonstrated enhanced robustness and accuracy in challenging traffic conditions.

Conclusions:

• The proposed algorithm effectively enhances lane detection and multi-object trajectory prediction capabilities.
• The integration of CA-HDC, lane fusion, trajectory adjustment, and S-GAN modules leads to superior performance over traditional methods.
• The algorithm shows significant potential for real-world autonomous driving systems and traffic analysis.