Improved Vehicle Object Detection Algorithm Based on Swin-YOLOv5s DOI Open Access
Haichao An, Jianhua Tang, Ying Fan

et al.

Processes, Journal Year: 2025, Volume and Issue: 13(3), P. 925 - 925

Published: March 20, 2025

In response to the challenges of low detection accuracy, slow speed, and high rates false positives missed detections in existing YOLOv5s vehicle models under complex traffic scenarios, an improved Swin-YOLOv5s algorithm is proposed this paper. By incorporating Swin Transformer attention mechanism replace original C3-1 network, computational load reduced capability capturing global features enhanced. The Self-Concat feature fusion method enhanced enable adaptive adjustment map weights, thereby enhancing positive features. results experiments conducted on KITTI dataset tests with Tesla V100 indicate that achieves a mean average precision (mAP) 95.7% F1 score 93.01%. These metrics represent improvements 1.6% 0.56%, respectively, compared YOLOv5s. Additionally, inference speed for single image increases by 1.11%, while overall frames per second (FPS) improves 12.5%. This enhancement effectively addresses issues related encountered severe occlusion conditions. ablation comparative different network validate both efficiency accuracy model, demonstrating its meet practical requirements more effectively.

Language: Английский

Improved Vehicle Object Detection Algorithm Based on Swin-YOLOv5s DOI Open Access
Haichao An, Jianhua Tang, Ying Fan

et al.

Processes, Journal Year: 2025, Volume and Issue: 13(3), P. 925 - 925

Published: March 20, 2025

In response to the challenges of low detection accuracy, slow speed, and high rates false positives missed detections in existing YOLOv5s vehicle models under complex traffic scenarios, an improved Swin-YOLOv5s algorithm is proposed this paper. By incorporating Swin Transformer attention mechanism replace original C3-1 network, computational load reduced capability capturing global features enhanced. The Self-Concat feature fusion method enhanced enable adaptive adjustment map weights, thereby enhancing positive features. results experiments conducted on KITTI dataset tests with Tesla V100 indicate that achieves a mean average precision (mAP) 95.7% F1 score 93.01%. These metrics represent improvements 1.6% 0.56%, respectively, compared YOLOv5s. Additionally, inference speed for single image increases by 1.11%, while overall frames per second (FPS) improves 12.5%. This enhancement effectively addresses issues related encountered severe occlusion conditions. ablation comparative different network validate both efficiency accuracy model, demonstrating its meet practical requirements more effectively.

Language: Английский

Citations

0