PEFNet: Position Enhancement Faster Network for Object Detection in Roadside Perception System

Roadside perception is a challenging research area that presents even greater difficulties than vehicle perception. Due to the different locations and angles of cameras, roadside objects exhibit violent multiscale variations, while vast sensing field introduces more small-scale targets complex backgrounds, making target recognition challenging. To address these problems, we focus on position information encoding achieve accurate object detection by proposing enhancement faster network (PEFNet). Based YOLOv6, FasterNet Block introduced into Backbone Neck networks provide efficient feature extraction achieving model lightweight transformation. improve small performance, position-aware pyramid (PA-PAN) proposed enhance encoding, SPD-Conv applied in PA-PAN further effective extraction. Finally, TSCODE integrated head suppress background noise interference. Experiments Rope3D UA-DETRAC datasets show our outperforms advanced YOLOX, FCOS detection. Compared with method improves mAP0.50 dataset from 78.18% 82.39%, AP such as pedestrians increasing 7.01%. Furthermore, PEFNet reduces weight 43.1% maintaining speed at 75fps higher accuracy previous algorithms for same number frames.