S2DB-mmWave YOLOv8n: Multi-object detection for millimeter-wave radar using YOLOv8n with optimized multi-scale...

Area of Science:

• Computer Vision
• Machine Learning
• Radar Technology

Background:

• Millimeter-wave (mmWave) radar offers all-weather, privacy-preserving object detection, crucial for intelligent security and transportation.
• Existing mmWave radar object detection faces challenges in distinguishing multiple targets and algorithmic performance.
• Deep learning approaches are vital for advancing mmWave radar capabilities.

Purpose of the Study:

• To propose an accurate deep learning-based target detection and classification framework for mmWave radar.
• To enhance feature extraction, detail recovery, and feature fusion in mmWave radar object detection.
• To address limitations in multi-target discrimination and detection performance.

Main Methods:

• Developed a novel backbone network with new convolutional layers and Simplified Spatial Pyramid Pooling - Fast (SimSPPF) module.
• Integrated a dynamic up-sampling technique for improved fine detail recovery.
• Incorporated a bidirectional feature pyramid network (BiFPN) for optimized feature fusion.

Main Results:

• The S2DB-mmWave YOLOv8n model achieved 93.1% mAP@0.5, 55.8% mAP@0.5:0.95, 89.4% precision, and 90.6% recall.
• Demonstrated significant improvements over the baseline YOLOv8n network (3.3%, 1.6%, 4.5%, and 7.7% higher, respectively).
• Performance gains were achieved without increasing the model's parameter count.

Conclusions:

• The proposed S2DB-mmWave YOLOv8n framework offers superior performance for mmWave radar object detection and classification.
• The novel architectural enhancements effectively address challenges in multi-target scenarios.
• This framework holds significant practical value for intelligent security and transportation applications.