Forest mobile robots based on visual navigation face the problem of limited computational power as edge devices and the navigation performance is greatly affected by illumination. To address this, a lightweight trunk detection method is proposed. This method uses visible and thermal image as inputs, minimizing the impact of illumination on navigation performance it also employs a feature extraction module based on Partial Convolution (PConv) and a Partial Efficient Layer Aggregation Network (P-ELAN) to achieve lightweight improvements to the baseline model. During training, the alpha-CioU loss function is used to replace the original CIoU loss function, increasing the accuracy of bounding box regression. The results show that the proposed tree trunk detection method for forest mobile robots reduces the parameter count of the original YOLOv7-tiny model by 31.7%, decreases computation by 33.3%, and improves inference speeds on Graphics Processing Units (GPU) and Central Processing Units (CPU) by 33.3% and 7.8%. The modified model maintains comparable accuracy while being more lightweight, making it an ideal choice for deployment on edge devices such as robots.
森林环境本质上是复杂且不可预测的[3]。如何进行实时的环境判别成为一个重要的主题。一种做法是采用激光雷达(Light Detection and Ranging,LiDAR)系统对周围环境进行实时扫描,并根据激光的往返时间进行精确距离测量。Malavazi等[4]通过点云处理激光雷达数据进行农作物和杂草检测,取得了不错的成果。然而,激光雷达提供的信息有限,因为其仅捕获距离和角度数据,不能从中对感知到的物体进行分类。另一种做法是使用相机作为导航系统的输入设备。相机具备成本低且易于安装的优势,同时能提供丰富的信息。基于此,相机在导航系统中得到了广泛的应用。Takagaki等[5]使用相机提供的颜色信息,以及阴影和土壤纹理,成功区分可通行区域(犁沟)和不可通行区域(山脊)。森林移动机器人全天候的工作内容对应着不同的光照条件,这对开发高效的机器人视觉系统提出重大挑战[6]。而普通RGB相机对光照敏感,弱光照下其拍照精度可能降低[7]并导致目标识别受损。与普通RGB相机不同,热成像仪具有不受光照条件阻碍的检测能力。Beyaz等[8]使用热成像仪检测树干的内部损坏,从而评估树干的健康情况。因此将热成像仪应用于森林移动机器人的导航是可行的。此外,人工智能和深度学习也被用于林业移动机器人的视觉导航。Itakura等[9]将YOLOv2和ResNet-50用于使用城市街道图像自动检测树木。Xie等[10]提出利用Faster R-CNN进行树木检测的注意力网络。
交并比(Intersection over Union,IoU,式中记为IoU)用于评估预测边界框和真实边界框之间的匹配程度。IoU仅考虑二者的位置重叠,而不考虑其位置和尺度偏差。因此,Zheng等[14]提出一种更精确的评估指标,称为完备交并比(Complete Intersection over Union,CIoU,式中记为LCIoU)。其表达式为
在本研究中,使用参数量和浮点运算数(Floating Point Operations,FLOPs)来评估模型大小;使用平均精度均值(Mean Average Precision,mAP,式中记为mAP)在不同IoU阈值(0.5~0.95,步长0.05)下的平均数即mAP@.5:.95来评估模型检测精度;使用单张图片推理时间来评估检测速度。mAP被认为是评估目标检测模型整体精度的重要指标,是其性能的可靠指标,计算公式为
YANGS, HONGT, ZHUL K.Forest fire prevention mobile robot path planning based on improved ant colony algorithm[J].Forest Engineering,2024,40(1):152-159.
[5]
BERGERMANM, BILLINGSLEYJ, REIDJ,et al.Robotics in agriculture and forestry[J].Springer Handbook of Robotics,2016:1463-1492.
[6]
MALAVAZIF B P, GUYONNEAUR, FASQUELJ B,et al.LiDAR-only based navigation algorithm for an autonomous agricultural robot[J].Computers and Electronics in Agriculture,2018,154:71–79.
[7]
TAKAGAKIA, MASUDAR, IIDAM,et al.Image processing for ridge/furrow discrimination for autonomous agricultural vehicles navigation[J].IFAC Proceedings Volumes,2013,46(18):47-51.
[8]
ZHANGY, CHENH, HEY,et al.Road segmentation for all-day outdoor robot navigation[J].Neurocomputing,2018,314:316-325.
[9]
GUOZ, LIX, XUQ,et al.Robust semantic segmentation based on RGB-thermal in variable lighting scenes[J].Measurement,2021,186:110176.
[10]
BEYAZA, ÖZKAYAM T.Canopy analysis and thermographic abnormalities determination possibilities of olive trees by using data mining algorithms[J].Notulae Botanicae Horti Agrobotanici Cluj-Napoca,2021,49(1):12139-12139.
[11]
ITAKURAK, HOSOIF.Automatic tree detection from three-dimensional images reconstructed from 360° spherical camera using YOLO v2[J].Remote Sensing,2020,12(6):988.
[12]
XIEQ, LID, YUZ,et al.Detecting trees in street images via deep learning with attention module[J].IEEE Transactions on Instrumentation and Measurement,2019,69(8):5395-5406.
[13]
DASILVAD Q, DOSSANTOSF N, SOUSAA J,et al.Visible and thermal image-based trunk detection with deep learning for forestry mobile robotics[J].Journal of Imaging,2021,7(9):176.
[14]
WANGC Y, BOCHKOVSKIYA, LIAOH Y M.YOLOv7:Trainable Bag-of-freebies Sets New State-of-the-art for Real-time Object Detectors[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.2023:7464-7475.
[15]
CHENJ,KAO S, HEH,et al.Run,Don′t Walk:Chasing Higher FLOPS for Faster Neural Networks[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.2023:12021-12031.
[16]
ZHENGZ, WANGP, REND,et al.Enhancing geometric factors in model learning and inference for object detection and instance segmentation[J].IEEE Transactions on Cybernetics,2021,52(8):8574-8586.
[17]
HEJ, ERFANIS, MAX,et al.Alpha-IoU:A family of power intersection over union losses for bounding box regression[J].Advances in Neural Information Processing Systems,2021,34:20230-20242.
[18]
LIUW, ANGUELOVD, ERHAND,et al.Ssd:Single Shot Multibox Detector[C]//Computer Vision-ECCV 2016:14th European Conference,Amsterdam,The Netherlands,October 11-14,2016,Proceedings,Part I 14.Springer International Publishing,2016:21-37.
[19]
RENS, HEK, GIRSHICKR,et al.Faster r-cnn:Towards real-time object detection with region proposal networks[J].Advances in neural information processing systems,2015,39(6):1137-1149.