Hybrid Deep Learning-Gaussian Process Network for Pedestrian Lane Detection in Unstructured Scenes

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Pedestrian lane detection is crucial for autonomous and assistive navigation systems.
• Existing methods struggle in unstructured environments lacking clear lane markers.
• Accurate lane detection is vital for user safety in navigation applications.

Purpose of the Study:

• To introduce a novel hybrid deep learning-Gaussian process (DL-GP) network for robust pedestrian lane detection.
• To address the challenge of detecting lanes on arbitrary surfaces without painted markers.
• To improve the safety and reliability of navigation systems in complex environments.

Main Methods:

• A hybrid DL-GP network combining a convolutional encoder-decoder with a hierarchical Gaussian Process classifier was developed.
• The network segments scene images into lane and background regions.
• A new dataset of 5000 images was created for training and evaluation.

Main Results:

• The proposed DL-GP network demonstrated significant performance improvements over existing methods on the new dataset.
• The network effectively segments pedestrian lanes in unstructured environments.
• It generates uncertainty maps crucial for assessing segmentation confidence and ensuring safety.

Conclusions:

• The hybrid DL-GP network offers a powerful and reliable solution for pedestrian lane detection in challenging environments.
• The developed dataset will accelerate research in this field, particularly for deep learning applications.
• The approach enhances navigation system safety through accurate lane identification and uncertainty estimation.