基于改进YOLO 11n的轻量级青贮玉米田间收获人体检测方法

Lightweight Silage Corn Field Personnel Detection with Improved YOLO 11n

摘要:

针对青贮玉米收获机作业过程中田间人员常被密集作物遮挡时可见光的目标检测效果受限,红外成像可感知田间环境中人体与周围作物温度的显著差异但分辨率较低,采用传统单一模态方法存在的精度不足及漏检率高的问题,提出了一种基于改进YOLO 11n的红外与可见光图像融合的青贮玉米田间人体检测方法,通过融合红外与可见光来提升田间被遮挡人体检测的鲁棒性和准确性。在主干网络中以轻量化的LWGA(Lightweight global-aware)模块替代原有的SPPF(Spatial pyramid pooling fast)结构,增强多尺度特征提取能力并降低模型复杂度;结合多层次的C3k2_TBFE模块与深层的C2PSA_biformer注意力机制,强化对复杂背景及遮挡目标的特征表达能力;在颈部网络中引入C3k2_TAB模块,提升多尺度语义融合效果及小目标检测性能。构建青贮玉米收获田间人体被遮挡图像的数据集,利用该数据集对改进后的模型进行训练和验证。实验结果显示,改进模型在红外与可见光数据集上mAP@ 50分别达到91.68%和88.32% ,较原始模型分别提升4.48、3.13个百分点,其模型参数量和浮点运算量分别降低4.26%与7.9%。改进后的模型可满足在田间人体实时检测需求且可长期适应户外工作,并构建双摄像头图像智能识别及多模态感知系统,保证高效、高精度实时检测,可为青贮玉米收获机作业过程中复杂环境下田间人体的实时检测提供理论技术支撑。

Abstract:

Aiming to address the limitations of visible-light-based object detection during silage corn harvesting, where field workers were often obscured by dense crops, and compensate for the low resolution of infrared imaging despite its ability to distinguish temperature differences between humans and surrounding vegetation, a detection method based on the fusion of infrared and visible images was proposed by using an improved YOLO 11n architecture. The original spatial pyramid pooling fast (SPPF) structure in the backbone network was replaced with a lightweight lightweight global-aware ( LWGA) module to enhance multi-scale feature extraction and reduce model complexity. Multi-level C3k2_TBFE modules and a deep C2PSA_biformer attention mechanism were incorporated to strengthen the representation of complex backgrounds and occluded targets, while a C3k2_TAB module was introduced in the neck network to improve multi-scale semantic fusion and small-object detection performance. A dataset containing images of occluded field personnel in cornfields was constructed, and the improved model was trained and validated using this dataset. Experimental results showed that detection accuracies of 91.68% and 88.32% were achieved on infrared and visible datasets, respectively, with improvements of 4.48 and 3.13 percentage points compared with that of the original model. The model’ s parameter count and computational complexity were reduced by 4.26% and 7.9% , respectively. The improved model was capable of meeting real-time detection requirements in field environments and supporting longterm outdoor operation, and a dual-camera intelligent recognition and multi-modal perception system was constructed to ensure efficient, accurate, and real-time detection, providing theoretical and technical support for personnel detection under complex conditions during silage corn harvesting.