应用改进优鲁模型对机载热成像中野生动物种类的识别方法

Identification of Wild Animal Species in Airborne Thermal Imaging Using Improved Youlu Model Algorithm

2019年,在研究区域黑龙江东北虎林园、吉林汪清国家级自然保护区、吉林珲春东北虎国家级自然保护区、内蒙古大兴安岭汗马国家级自然保护区内,利用M300RTK型无人机并配套H20T红外热成像机,重点监测了东北虎及其主要猎物(马鹿、狍、梅花鹿、驯鹿)的生态行为,获取不同季节的野生动物影像,构建了野生动物监测影像数据库,其中,东北虎(421幅)、马鹿(378幅)、狍(419幅)、梅花鹿(381幅)、驯鹿(401幅),共计热成像图像2 000幅;采用优鲁V5s(YOLO V5s)模型对机载热成像中野生动物种类进行识别,分析改进前、后优鲁V5s模型识别效果的差异。实验结果表明:采用优鲁V5s模型算法对机载热成像中野生动物种类进行识别,准确率为94.1%,模型权重为14.8 MB。采用优化后的优鲁V5s模型结构对机载热成像中野生动物种类进行识别,准确率为93.2%,模型权重降为7.7 MB,在识别准确率仅下降0.9%的情况下,模型权重减小48%,单张图像检测时间从0.032 s降低到0.015 s,减少53%;改进后的模型有效降低了算法依托硬件的需求,可为无人机载系统野生动物在线调查工作提供一种轻量化边缘计算方法。

In 2019, in the study area of Siberian Tiger Park in Heilongjiang Province, Wangqing National Nature Reserve in Jilin Province, Hunchun Amur tiger National Nature Reserve in Jilin Province and Hanma National Nature Reserve in Daxing’an Mountains, Inner Mongolia, m300 rtk UAV and supporting h20 t infrared thermal imager were used to monitor the ecological behavior of Amur tiger and its main prey(red deer, roe deer, sika deer and reindeer) and obtain wildlife images in different seasons. The database of wildlife monitoring images-Amur tiger(421), red deer(378), roe deer(419), sika deer(381) and reindeer(401), with a total of 2 000 thermal imaging images;Youlu v5 s model was used to identify the species of wild animals in airborne thermal imaging, and the difference between the recognition effect of Youlu v5 s model before and after improvement was analyzed. The recognition accuracy of wild animal species in airborne thermal imaging is 94.1% and the weight of the model was 14.8 MB. The optimized Youlu v5 s model structure was used to identify wild animal species in airborne thermal imaging. The accuracy was 93.2%, and the model weight was reduced to 7.7 MB. When the recognition accuracy is only reduced by 0.9%, the model weight is reduced by 48%, and the detection time of a single image is reduced from 0.032 to 0.015 s, which is reduced by 53%. The improved model effectively reduces the hardware requirements of the algorithm, and provides a lightweight edge calculation method for the on-line wildlife investigation of unmanned airborne system.