Adaptive Fusion of LiDAR Features for 3D Object Detection in Autonomous Driving.

Area of Science:

• Autonomous Driving
• Sensor Fusion
• Computer Vision

Background:

• Cooperative perception in autonomous driving uses multi-sensor data (LiDAR, cameras, radar) for enhanced environmental understanding.
• Traditional early/late fusion methods struggle with bandwidth and computational costs, impacting perception accuracy, especially for small objects like pedestrians.
• Detecting small objects and achieving efficient, accurate cooperative perception remains a challenge in complex traffic scenarios.

Purpose of the Study:

• To propose a novel cooperative perception framework using two-stage intermediate-level sensor feature fusion.
• To address the limitations of existing fusion methods in balancing efficiency and accuracy for detecting pedestrians and vehicles.
• To enhance the perception capabilities in complex traffic environments with coexisting pedestrians and vehicles.

Main Methods:

• Developed a novel cooperative perception framework utilizing two-stage intermediate-level sensor feature fusion.
• Designed the framework for complex traffic scenarios involving both pedestrians and vehicles.
• Validated the model through qualitative and quantitative experiments on simulated and real-world datasets.

Main Results:

• The proposed framework demonstrates superior performance in detecting small objects like pedestrians compared to mainstream methods.
• Achieved improved cooperative perception accuracy for medium and large objects, such as vehicles.
• Outperformed state-of-the-art models, with up to 4.1% improvement in vehicle detection and 29.2% in pedestrian detection accuracy (mAP).

Conclusions:

• The novel two-stage intermediate-level sensor feature fusion framework effectively enhances cooperative perception in autonomous driving.
• The approach significantly improves the detection of small objects (pedestrians) and overall perception accuracy in complex traffic scenarios.
• Experimental validation confirms the reliability and superior performance of the proposed method over existing state-of-the-art models.