LiteMapNet: Lightweight Online High-Definition Mapping With Attention Head Pruning

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• High-definition (HD) maps are crucial for autonomous driving perception and planning.
• Current online HD map construction methods are computationally intensive, limiting real-time application on embedded systems.

Purpose of the Study:

• To develop LiteMapNet, a computationally efficient transformer-based framework for accelerated online HD map construction.
• To address the real-time deployment challenges of HD mapping on resource-constrained platforms.

Main Methods:

• Proposed LiteMapNet, a lightweight transformer framework utilizing structured attention-head pruning.
• Employed a two-stage pruning pipeline: Fisher information matrix for importance estimation and learnable masks for performance recovery.
• Utilized lightweight calibration to enhance stability and generalization without inference overhead.

Main Results:

• LiteMapNet significantly reduces computational cost and floating-point operations.
• Achieved up to 1.5x inference speedup on NVIDIA H100 GPUs.
• Maintained mapping accuracy with less than 1% drop compared to the unpruned baseline.

Conclusions:

• LiteMapNet enables efficient online HD map construction for autonomous driving.
• The framework facilitates deployment on resource-limited in-vehicle platforms.
• The proposed pruning and calibration strategy balances efficiency and accuracy effectively.