摘要
针对现代化养殖业无人化、智能化的需求,以目标检测网络YOLOv2为基础,提出了一种基于深度学习提取时空特征的生猪动作识别与定位的方法。对待检测视频关键帧中的生猪空间位置信息与视频流时序动作特征进行检测,采用通道注意力模块将这2种特征进行合理且平滑的特征融合,实现了一个端到端的动作识别网络,可以直接从视频序列中预测得到关键帧的包围框和动作分类概率。通过对某生猪养殖场群养栏监控视频进行训练和测试,研究了通道注意力模块和网络输入视频帧采样间隔对检测效果的影响,验证了三维卷积神经网络在生猪动作识别与定位中的有效性。
In order to meet the unmanned and intelligent needs in modern breeding industry,based on the object detection network yolov2,we propose a method for pig action recognition and localization,using deep learning to extract temporal and spatial features.In order to detect the spatial position information of pigs in the key video frames and the temporal motion features of video stream,the channel attention module is used to fuse these two features reasonably and smoothly,and an end-to-end action recognition network is realized.The bounding frame and action classification probability of key frames can be predicted directly from the video sequence.Through training and testing the monitoring video of a pig farm,the effect of channel attention module and network input video frame sampling interval on the detection effect is studied,verifying the effectiveness of 3D-convolutional neural networks(CNN)in pig action recognition and positioning.
作者
苏森
陈春雨
刘文龙
李诚
SU Sen;CHEN Chunyu;LIU Wenlong;LI Cheng(College of Information and Communication Engineering,Harbin Engineering University,Harbin 150001,China)
出处
《应用科技》
CAS
2021年第4期80-84,共5页
Applied Science and Technology
基金
国家自然科学基金项目(61871142)
基于人工智能架构的多传感器信息融合与决策系统的研究与实现(KY10800180032)
中央高校基本科研业务费项目(3072020CFT0803).
