Lightweight recognition network for dense crowd images based on YOLO11
TL;DR
The model adopts MobileViT as its backbone network, which effectively enhances the overall feature extraction capability of the model in dense crowd scenarios and further aggregates pixel-level features through dimensional interaction and improves small-object detection by integrating a 160×160-scale detection head.
Abstract
In response to the challenges faced by autonomous driving and crowd flow monitoring systems in dense occlusion crowd detection, as well as the low accuracy and high computational cost of other detection algorithms, an improved lightweight dense crowd detection model named MGN-YOLO based on YOLO11 is proposed. The model adopts MobileViT as its backbone network, which effectively enhances the overall feature extraction capability of the model in dense crowd scenarios. To encode global information, the GMA attention mechanism module is incorporated, which further aggregates pixel-level features through dimensional interaction and improves small-object detection by integrating a 160×160-scale detection head. By employing CIoU-NWD as the bounding box loss function, the issues of missed detection and false detection for small-scale crowded objects in dense scenarios are effectively alleviated. The experimental results show that compared with YOLO11n, the accuracy of the MGN-YOLO crowd detection algorithm is improved by 4.4%(mAP_0.5/%) on the CrowdHuman and 2.1% on the WiderPerson, while only 2.7M parameters and 7.7 GFLOPs of model calculations, which meet the deployment requirements of low computing power and high precision.