Enhancing YOLOv5 for Autonomous Driving: Efficient Attention-Based Object Detection on Edge Devices.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Autonomous Systems

Background:

• Object detection is crucial for autonomous driving safety and efficiency.
• Deep learning models are effective but computationally expensive for edge devices.
• There is a need for lightweight, high-performance object detection models.

Purpose of the Study:

• To develop lightweight object detection models for real-time autonomous driving on edge devices.
• To integrate advanced channel attention strategies (ECA, SE) into the YOLOv5s architecture.
• To evaluate model performance on standard datasets like KITTI and BDD-100K.

Main Methods:

• Utilized the YOLOv5s architecture as a base for lightweight object detection.
• Integrated Efficient Channel Attention (ECA) and Squeeze-and-Excitation (SE) attention modules.
• Trained and evaluated four distinct models on the KITTI and BDD-100K datasets.
• Assessed performance using metrics such as precision, recall, and mean average precision (mAP).

Main Results:

• BaseECAx2 model achieved the lowest GFLOPs (13) and smallest size (9.1 MB), ideal for edge devices.
• BaseSE-ECA model demonstrated high accuracy with 96.69% precision and 98.4% mAP for vehicle detection.
• Models showed reduced performance in challenging conditions (low-light, motion blur) on the BDD-100K dataset.

Conclusions:

• Lightweight YOLOv5s models with attention mechanisms offer a balance of performance and efficiency for autonomous driving.
• The BaseECAx2 and BaseSE-ECA models present cost-effective solutions for real-time edge deployment.
• Further research is needed to improve robustness in complex, real-world driving scenarios.