In recent years, road traffic object detection has gained prominence in areas such as traffic monitoring, autonomous driving, and road safety. Nonetheless, existing algorithms offer room for improvement, particularly when detecting distant or inherently small targets, such as vehicles and pedestrians, from camera perspectives. By addressing the detection accuracy issues associated with small targets, this study introduces the YOLOv5s-LGC detection algorithm. This model incorporates a multiscale feature fusion network and leverages the lightweight GhostNet module to reduce model parameters. Furthermore, the GC attention module is employed to mitigate background interference, thereby enhancing the average detection accuracy across all categories. Through data analysis, target detection at different scales and sampling rates is determined. Experiments indicate that the YOLOv5s-LGC model surpasses the baseline YOLOv5s in detection accuracy on the Partial_BDD100K and KITTI datasets by 3.3% and 1.6%, respectively. This improvement in locating and classifying small targets presents a novel approach for applying object detection algorithms in road traffic scenarios.
Research on Small Target Detection Algorithm for Autonomous Vehicle Scenarios
2025
Aufsatz (Zeitschrift)
Elektronische Ressource
Unbekannt
Metadata by DOAJ is licensed under CC BY-SA 1.0