Sec-CLOCs: Multimodal Back-End Fusion-Based Object Detection Algorithm in Snowy Scenes.

Area of Science:

• Computer Vision
• Autonomous Driving Systems
• Sensor Fusion

Background:

• Adverse weather conditions like heavy snow significantly degrade LiDAR and camera performance in intelligent driving vehicles.
• Current environmental sensing capabilities are insufficient for safe operation under severe weather, posing risks to driving safety.

Purpose of the Study:

• To propose and evaluate Sec-CLOCs, a multimodal back-end fusion object detection method optimized for vehicle detection in heavy snowfall.
• To enhance the environmental sensing capabilities of autonomous vehicles under challenging weather conditions.

Main Methods:

• Integration of an improved YOLOv8s 2D detector with a SECOND 3D detector using a back-end fusion approach.
• Image data enhancement via the Two-stage Knowledge Learning and Multi-contrastive Regularization (TKLMR) algorithm.
• Point cloud data preprocessing using the LIDROR algorithm for 3D detection, followed by CLOCs fusion of 2D and 3D results.

Main Results:

• Sec-CLOCs achieved 82.34% vehicle detection accuracy in moderate conditions (30-100 m) and 81.76% in hard conditions (>100 m) during heavy snowfall.
• The method demonstrates high detection performance and robustness in severe snow environments.
• Optimizations included DyHead detection head and Wise-IOU loss function for YOLOv8s, and TKLMR for image quality.

Conclusions:

• The Sec-CLOCs algorithm significantly improves vehicle detection performance in heavy snowfall conditions.
• Multimodal back-end fusion of enhanced 2D and 3D sensor data offers a robust solution for autonomous driving safety in adverse weather.
• The proposed method enhances the reliability of environmental sensing for intelligent vehicles.