基于红外热成像与改进YOLOV3的夜间野兔监测方法

Night hare detection method based on infrared thermal imaging and improved YOLOV3

随生态改善,野兔数量增多,对农田与林地的危害日益加重。野兔活动多为夜间,目标小,运动速度快,且出现环境较复杂,监控兔害,需要一种高效智能化的方法。针对野兔活动习性,该文提出了使用红外热成像实时监控,结合改进的YOLOV3目标检测方法对夜间野兔进行检测。根据YOLOV3目标检测网络基本结构提出了一种针对红外图像中野兔的实时检测的网络(infrared rabbit detection YOLO,IR-YOLO),该网络特征提取部分压缩YOLOV3特征提取网络深度,利用浅层卷积层特征以提高低分辨率红外小目标检测精度,降低运算量,网络检测部分使用基于CenterNet结构的检测方式以提高检测速度。使用热成像野外实时采集的夜间野兔图像作为数据集,包括不同距离,尺度,出现环境不同的野兔共计6000幅红外图像制作训练集与测试集,比例为5:1。试验结果表明,IR-YOLO在红外热成像视频中复杂环境下出现的野兔检测率达75%,平均检测速度51帧/s,相对改进前YOLOV3检测率提高15个百分点,相对改进前YOLOV3检测速度提高5帧/s。相比其他目标检测算法各项检测指标更为优良,检测率方面相对Faster-RCNN与RFCN-RESNET101分别提高45个百分点与20个百分点,检测速度方面相对Faster-RCNN与RFCN-RESNET101分别提高30和与45帧/s。该方法可高效快速地对夜间复杂环境下出现的野兔进行检测,也可广泛应用于夜间对其他类型农业害兽的检测。

Restoration of and improvement in ecosystems has led to a growth in hare numbers to the detriment of farmland and woodland where they appear.Understanding forage of the hare is imperative to better manage them but proves difficult because most hares forage fruit forests and mountain farmlands in night.In this paper we present a method to monitor real-time forage of the hares using infrared thermal imaging and the improved YOLOV3 target detection method-the real-time infrared hare detection network-YOLO(IR-YOLO)network.The feature-extraction component in the proposed network compressed the depth of the feature-extract in the YOLOV3 in attempts to improve the accuracy of detecting small targets from low-resolution infrared images by using the shallow-layer convolution method.In order to simplify computation and improve detection efficiency and accuracy of small targets,the detection component in the network was based on the Center Net structure.Considering that hares are more active in autumn,real-time hare forage in the infrared images taken from wheat and rape fields were used to validate the proposed method.The datasets include 6000 infrared images taken at different scales and environments,and the ratio of the datasets used for training to the sets used for validating was 5:1.The results show that the proposed method identified 75%of hares based on the infrared thermal image videos.The average processing speed of the proposed method is 51 frames/s,the detection rate is15 percentage points higher than that of YOLOV3,Compared with the Faster-RCNN and RFCN-RESNET101,the proposed method increased processing speed by 45 and 20 percentage points respectively.In summary,the proposed method can quickly and accurately detect the forage of hares in complex environment in night,and it can be used to monitor night-movement of other animals appearing in other ecosystems.