【摘要】 青贮玉米田复杂环境遮挡驾驶员视线，收获机作业时易造成人员伤亡和重大经济损失。针对青贮收获机频繁伤害人畜的安全问题，构建了轻量化热红外图像检测模型YOLO-TI,设计了适用于青贮收获机的热红外活体检测预警系统。采集不同时间、地点和姿态的活体热红外图像并进行预处理，优化YOLO v5模型。将YOLO v5的主干网络替换为SqueezeNet进行特征提取，融入红外空间注意力机制(ISAM)以增强目标特征，并将损失计算中的CIoU函数替换为SIoU,构建YOLO-TI模型。试验结果显示YOLO-TI检测精确率达99.6%、召回率为99.1%、F1值为99.0%,实时检测速度为48 f/s。对于相同的测试数据集，提出的模型较YOLO v5模型精确率、召回率、F1值以及检测速度分别提升4.4、2.1、3.3个百分点和11 f/s;与YOLO v10相比，其精确率和召回率分别提升0.1、1.4个百分点。田间试验结果表明，在18℃环境下，系统在作业范围0～15 m内活体检测精度平均值达91.5%,在收获机最小安全距离2.32 m内，系统从目标出现到报警激活响应时间为0.45 s,符合安全收获作业要求。该系统满足作业过程中实时检测活体的需求，且安全可靠，为青贮收获机在复杂农业场景下精确、高效检测活体提供了技术支持。更多还原

【Abstract】 In silage maize fields, the complex environment often obstructs the driver’s view, making harvesting operations prone to causing casualties and severe economic losses. To address the frequent accidents involving silage harvesters injuring humans and animals, a lightweight thermal infrared image detection model, YOLO-TI, was developed, and a thermal infrared-based living-body detection and early-warning system suitable for silage harvesters was designed. Thermal infrared images of living bodies under different times, locations, and postures were collected and preprocessed to optimize the YOLO v5 model. The backbone network of YOLO v5 was replaced with SqueezeNet for feature extraction, while an infrared spatial attention mechanism(ISAM) incorporated to enhance target features. Additionally, the CIoU loss function was replaced with SIoU to construct the YOLO-TI model. Experimental results showed that YOLO-TI achieved detection precision of 99.6%, recall of 99.1%, F1 score of 99.0%, and real-time detection speed of 48 f/s. On the same test dataset, compared with YOLO v5, the proposed model improved precision, recall, F1 score, and detection speed by 4.4, 2.1, 3.3 percentage points, and 11 f/s, respectively; compared with YOLO v10, precision and recall were improved by 0.1, 1.4 percentage points, respectively. Field test results indicated that at ambient temperature of 18℃, the system achieved average detection accuracy of 91.5% within operating range of 0～15 m. Within the harvester’s minimum safety range of 2.32 m, the response time from target appearance to alarm activation was 0.45 s, meeting the safety requirements of harvesting operations. The system satisfied the demand for real-time detection of living bodies during operation, ensuring safety and reliability, and provided technical support for accurate and efficient living-body detection in complex agricultural scenarios with silage harvesters.更多还原