CenterPNets: A Multi-Task Shared Network for Traffic Perception

Area of Science:

• Computer Vision
• Autonomous Driving Systems
• Deep Learning Architectures

Background:

• Panoramic traffic perception is crucial for autonomous driving safety and efficiency.
• Existing methods often require separate models for distinct perception tasks, limiting resource utilization and overall performance.
• High-accuracy shared networks are needed to integrate multiple traffic sensing tasks.

Purpose of the Study:

• To introduce CenterPNets, a unified multi-task shared sensing network for autonomous driving.
• To enhance the accuracy and efficiency of simultaneous target detection, drivable area segmentation, and lane detection.
• To propose key optimizations for improved multi-task learning performance.

Main Methods:

• Developed CenterPNets, a multi-task shared sensing network integrating detection, segmentation, and lane detection heads.
• Utilized a shared path aggregation network for efficient feature reuse between tasks.
• Implemented an anchor-free detection head for faster target localization and a split-head branch for multi-scale feature fusion.
• Designed an efficient multi-task joint training loss function for model optimization.

Main Results:

• CenterPNets achieved 75.8% average detection accuracy on the Berkeley DeepDrive dataset.
• The network demonstrated high performance in segmentation tasks with 92.8% intersection ratio for drivable areas and 32.1% for lane areas.
• Optimizations led to improved overall detection performance and inference speed.

Conclusions:

• CenterPNets provides a precise and effective solution for integrated multi-task traffic perception.
• The proposed network architecture and optimizations significantly advance the capabilities of autonomous driving systems.
• This unified approach demonstrates the potential of shared networks for complex real-world traffic sensing challenges.