Deep Inference Networks for Reliable Vehicle Lateral Position Estimation in Congested Urban Environments

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Accurate vehicle lateral position estimation is crucial for autonomous vehicle safety.
• Road occlusion and unreliable reference objects present significant challenges for existing methods.

Purpose of the Study:

• To propose a novel deep inference network (DINet) for robust vehicle lateral position estimation.
• To address limitations of current approaches in handling road occlusion and unreliable reference objects.

Main Methods:

• DINet integrates three deep neural network (DNN) components: RADOOS for road area detection and object segmentation, RAR for road area reconstruction, and LPE for lateral position estimation.
• The RAR model infers missing road regions conditioned on segmented occluding objects.
• The LPE model estimates lateral position from the reconstructed road area.

Main Results:

• DINet demonstrated reliable and accurate (centimeter-level) lateral position estimation in road tests.
• The system performed effectively even under severe road occlusion scenarios.
• Experimental results validate the proposed network's capability to overcome common challenges.

Conclusions:

• The proposed DINet effectively estimates vehicle lateral position, enhancing autonomous driving safety.
• DINet offers a robust solution for scenarios with significant road occlusion.
• The human-like integration of DNN components enables realistic road area reconstruction and accurate positioning.