Semantic Scene Completion in Autonomous Driving: A Two-Stream Multi-Vehicle Collaboration Approach

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Vehicle-to-vehicle (V2V) communication aids autonomous driving by sharing sensor data.
• Feature misalignment between vehicles causes ambiguity and reduces semantic scene completion accuracy.

Purpose of the Study:

• To propose a novel approach for collaborative semantic scene completion in autonomous driving.
• To address the challenges of feature misalignment in V2V communication.

Main Methods:

• Introduced a Two-Stream Multi-Vehicle (TSMV) approach, processing collaborative features in two streams.
• Developed the Neighborhood Self-Cross Attention Transformer (NSCAT) module to query similar local features without assuming synchronization.
• Generated a 3D occupancy map from aggregated collaborative vehicle features.

Main Results:

• The TSMV approach demonstrated superior performance compared to existing methods.
• Experiments were conducted on V2VSSC and SemanticOPV2V datasets.
• The NSCAT module effectively mitigated issues arising from feature misalignment.

Conclusions:

• TSMV significantly improves collaborative semantic scene completion accuracy.
• The proposed method offers a robust solution for V2V-based autonomous driving perception.
• Future work can explore further enhancements in multi-vehicle collaboration for complex driving scenarios.