Modified YOLO11 for improving car detection performance on the highway

Abstract

Vehicle detection in the road area serves as the basis for various other studies, including vehicle type identification, vehicle counting, and traffic violation detection. Two models that have been successfully implemented were YOLOv2 and YOLOv3, but both still have limitations in detecting small objects and require significant computational resources. YOLO11 displays enhanced accuracy and computational efficiency. Large variant models such as YOLO11m, YOLO11l, and YOLO11x achieve higher accuracy but often come with a substantial decrease in inference speed. Therefore, this study proposes modifying the YOLO11 architecture by removing two convolutional blocks and three C3k2 Blocks to increase detection speed without sacrificing accuracy. The proposed method is applied to two datasets: low-traffic and high-traffic highways. The test results showed that the suggested approach could reduce the inference time by 19% on low-traffic highways and 12% on high-traffic highways, respectively. In addition, the proposed method can increase the Average Precision (AP) by 0.004 on low-traffic highways and 0.003 on high-traffic highways, making it better suited for real-time car detection on highways.