基于改进轻量深度网络的牧区牲畜目标快速检测  

Rapid detection of livestock targets in pastoral areas based on improved lightweight deep network

作　　者：朱俊峰[1,2,3] 刘洋[3] 王星天 曹亮 ZHU Junfeng;LIU Yang;WANG Xingtian;CAO Liang(Yinshanbeilu Grassland Eco-hydrology National Observation and Research Station,China Institute of Water Resources and Hydropower Research,Beijing 100083,China;Institute of Water Resources for Pastoral Area,Ministry of Water Resources,Hohhot 010020,China;Inner Mongolia Key Laboratory of Mechanical&Electrical Control,Hohhot 010020,China)

机构地区：[1]中国水利水电科学研究院内蒙古阴山北麓草原生态水文国家野外科学观测研究站,北京100083 [2]水利部牧区水利科学研究所,呼和浩特010020 [3]内蒙古机电控制重点实验室,呼和浩特010020

出　　处：《生态学杂志》2023年第11期2786-2796,共11页Chinese Journal of Ecology

基　　金：内蒙古机电控制重点实验室开放基金项目(IMMEC2020001);国家重点研究发展计划项目(2018YFE0196000)资助。

摘　　要：为实现农牧区牲畜目标的快速、准确检测,提出一种改进YOLOV3⁃tiny的轻量级牧区牲畜目标检测算法,并在Jetson Nano嵌入式主板上实现实时检测。该算法首先根据牧区牲畜体型相差较大的特点优化了网络结构,引入一种锚框复合聚类算法,并增加预测输出尺度,增强浅层信息的利用;其次,采用金字塔网络进行多尺度特征融合,在保证大目标检测率的同时提高小目标检测率;最后针对复杂光照条件下(如太阳光直射下)检测精度下降问题,加入注意力机制,提高复杂光照条件下目标检测精度。实验结果表明:改进后YOLOV3⁃tiny算法检测精度达83.2%,在嵌入式平台Jetson Nano主板上的检测速度为12帧·s^(-1),相较于YOLOV3⁃tiny算法平均检测精度提高了8.7%。To achieve fast and accurate detection of livestock targets in grazing areas,we proposed a lightweight livestock target detection algorithm with improved YOLOV3⁃tiny,which is a lightweight object detection algorithm for real⁃time detection on Jetson Nano embedded motherboard.In terms of network structure,the anchor frame clus⁃tering algorithm is optimized according to the characteristics of livestock targets in grazing areas,and the prediction output scale is increased to enhance the use of shallow information.The pyramid network is used for multi⁃scale fea⁃ture fusion to improve the detection rate of small targets while ensuring the detection rate of large targets.The im⁃proved target detection mechanism can effectively improve the accuracy of target detection under complex light con⁃ditions(e.g.,direct sunlight).The experimental results showed that the detection accuracy of the improved YOLOV3⁃tiny algorithm reached 83.2%,and the detection speed on the embedded platform Jetson Nano was 12 frames·s^(-1).The algorithm improved the detection accuracy by 8.7%on average compared with the YOLOV3⁃tiny algorithm while satisfying the portability.

关 键 词：卷积神经网络 YOLO网络 深度学习 智慧牧区 

分 类 号：S818.9[农业科学—畜牧学] TP391.41[农业科学—畜牧兽医]