欢迎访问林业科学,今天是

林业科学 ›› 2026, Vol. 62 ›› Issue (11): 140-152.doi: 10.11707/j.1001-7488.LYKX20260378

• 研究论文 • 上一篇    

基于YOLO-MAN的西藏棕熊检测模型与智能驱离原型系统

沈忱1,杨雅萱2,翟波3,谢将剑1,4,*(),谢屹3   

  1. 1. 北京林业大学工学院 北京 100083
    2. 中央民族大学信息工程学院 北京 100089
    3. 北京林业大学经济管理学院 北京 100083
    4. 北京林业大学林木资源高效生产全国重点实验室 北京 100083
  • 收稿日期:2026-06-17 修回日期:2026-08-20 出版日期:2026-11-10 发布日期:2026-11-20
  • 通讯作者: 谢将剑 E-mail:shyneforce@bjfu.edu.cn
  • 基金资助:
    北京林业大学科技创新计划(2024XY?G002)。

Detection Model and Intelligent Deterrence Prototype System for Tibetan Brown Bears Based on YOLO-MAN

Chen Shen1,Yaxuan Yang2,Bo Zhai3,Jiangjian Xie1,4,*(),Yi Xie3   

  1. 1. School of Technology, Beijing Forestry University Beijing 100083
    2. School of Information Engineering, Minzu University of China Beijing 100089
    3. School of Economics and Management, Beijing Forestry University Beijing 100083
    4. State Key Laboratory of Efficient Production of Forest Resources, Beijing Forestry University Beijing 100083
  • Received:2026-06-17 Revised:2026-08-20 Online:2026-11-10 Published:2026-11-20
  • Contact: Jiangjian Xie E-mail:shyneforce@bjfu.edu.cn

摘要:

目的: 针对三江源地区日益突出的人熊冲突问题,开展面向牧区场景的西藏棕熊自动检测与非致命驱离技术研究,以降低人工近距离驱赶风险,提高牧区人熊冲突的安全防护能力,为三江源地区人熊冲突预警与非致命干预提供技术参考。方法: 设计一种由西藏棕熊检测模块、云台跟踪模块和主动驱离模块组成的智能驱离原型系统。利用本研究在YOLO11基础上面向西藏棕熊检测任务适配构建的YOLO-MAN模型对目标进行检测与定位;系统根据目标偏移量控制云台持续跟踪,并在目标进入预设范围后触发电动脉冲水枪喷洒辣椒素溶液,实现非致命定向驱离。引入并适配混合聚合网络(MANet)和全局特征聚合结构,协同构建多尺度特征增强策略,以提高模型对复杂环境下西藏棕熊的特征表达与定位能力;同时,通过引入三江源同域分布动物以及跨域相似动物类别作为干扰样本,构建抗干扰负样本训练方式,增强模型的抗干扰判别能力。在此基础上,构建面向三江源场景的复合野生动物检测数据集,并将模型部署于Jetson Nano B01边缘计算平台。结果: YOLO-MAN在测试集上的P、R、mAP@0.5和mAP@0.5:0.95分别达到88.3%、86.6%、91.0%和75.2%,其中mAP@0.5:0.95较YOLO11提升2.8%,模型的参数量和计算量分别为3 237 589与8.2 GFLOPs。消融试验表明,MANet与全局特征聚合结构均能够有效提升模型检测性能,mAP@0.5:0.95分别提升1.7%和1.1%,二者结合后效果最佳,提升了2.8%。边缘部署结果表明,经TensorRT加速后模型在 Jetson Nano B01平台上的推理速度约为9帧·s–1,能够满足原型系统的实时检测需求。结论: 构建的YOLO-MAN模型能够提升复杂野外场景下西藏棕熊的检测与定位性能,所构建的智能驱离原型系统初步验证了目标检测、云台跟踪与主动非致命驱离相结合技术路线的可行性。

关键词: 人熊冲突, 非致命驱离, 西藏棕熊监测, 混合聚合网络, 全局特征聚合结构协同

Abstract:

Objective: In view of the growing human-bear conflict in the Sanjiangyuan region, this study aims to explore the technology for automated detection and non-lethal deterrence of Tibetan brown bears in pastoral areas, in order to reduce the risks of close-range manual intervention, improve the security protection ability of human bear conflict in the pastoral area, and provide technical reference for early warning and non-lethal intervention of human bear conflict in the Sanjiangyuan region. Method: An intelligent deterrence prototype system consisting of a Tibetan brown bear detection module, a gimbal tracking module, and an active deterrence module was designed. A YOLO-MAN model built on the basis of YOLO11 to adapt to the Tibetan brown bear detection task was used for real-time detection and localization of the bear. The system controlled the gimbal to continuously track the target based on the target offset, and triggered the electric arterial flushing gun to spray capsaicin solution when the target entered the predefined range, achieving non-lethal directional drive away. The mixed aggregation network (MANet) and a global feature aggregation module were introduced to collaboratively build multi-scale feature enhancement strategy to improve the model's ability to express and locate features of Tibetan brown bears in complex environments. At the same time, by introducing animals with the same distribution in the Sanjiangyuan region and similar animal categories (yaks, bharals, giant pandas, and red pandas) across domains as interference samples, an anti-interference negative sample training method was constructed to enhance the model's anti-interference discrimination ability. On this basis, a composite wildlife detection dataset for the Sanjiangyuan scenario was constructed, and the model was deployed on a Jetson Nano B01 edge computing platform. Result: YOLO-MAN’s P, R mAP@0.5 and mAP@0.5:0.95 on the test set achieved 88.3%, 86.6%, 91.0%, and 75.2%, respectively, among which mAP@0.5 0.95 had a 2.8% increase compared to YOLO11. The parameter count and computational complexity of the model were 3 237 589 and 8.2 GFLOPs, respectively. Ablation experiments showed that MANet and global feature aggregation structures were able to effectively improve the model detection performance, and mAP@0.5 0.95 increased by 1.7% and 1.1%, respectively, and the combination of the two resulted in the best effect, with an increase of 2.8%. The results of edge deployment showed that after TensorRT acceleration, the model reached approximately 9 frames per second on the Jetson Nano B01, meeting the prototype’s real-time requirements. Conclusion: YOLO-MAN model built in this study can improve the detection and localization performance of Tibetan brown bears in complex field environments. The developed intelligent deterrence prototype system preliminarily demonstrates the feasibility of integrating target detection, pan-tilt tracking, and active non-lethal deterrence.

Key words: human-bear conflict, non-lethal deterrence, Tibetan brown bear monitoring, mixed aggregation network, collaborative global feature aggregation structure

中图分类号: