GY-YOLO: ghost separable YOLO for pedestrian detection
Ali M. Elhenidy, Labib M. Labib, Amira Y. Haikal, Mahmoud M. Saafan
Abstract
Abstract In recent years, there has been impressive development in human detection. The main challenge in pedestrian detection is the training data. To assess detectors in crowd scenarios more effectively, a novel dataset in this study called the HEP dataset (Hybrid Egyptian Pedestrian dataset) is introduced. The HEP dataset is extensive, has comprehensive annotations, and is highly diverse. The dataset images are collected by two different means. Most of the images are collected from different mobile cameras for people crossing the street in high crowded streets in Egypt, and the rest of the images are collected from the web. That is why the dataset is called hybrid. The collected dataset is more suitable for pedestrian detection as the whole images focus on pedestrian scenarios for people outdoors crossing the street. This outperforms the previous benchmarks such as CrowdHuman and WiderPerson which collect data from the web and surveillance cameras with lots of images for indoor people. GS-YOLO also is proposed to address the real-time performance and the occlusion in the crowd scenes issues. GS-YOLO is a novel pedestrian detection model that utilizes efficient Ghost and depth separable convolution modules. GS-YOLO replaces all the convolution layers in the backbone and the head of the original YOLOv8 with Ghost and depth separable modules, respectively. A deformable to-features module is proposed to enrich features for the different feature pyramid networks. GS-YOLO is trained and tested over the collected dataset and other benchmarks like CrowdHuman and WiderPerson datasets. GS-YOLO achieves competitive results over the state-of-the-art models such as YOLOv5 and YOLOv8. GS-YOLO achieves 92.8% mAp on the HEP dataset, while YOLOv5 achieves 90.3% mAp and YOLOv8 achieves 91.1% mAp.