Research on Object Recognition in Warehouse Environment Based on Improved Faster R-CNN
扫码查看
点击上方二维码区域，可以放大扫码查看
原文链接
国家科技期刊平台
NETL
NSTL
万方数据
为解决传统目标检测精确度不高、有效性差、难以适应仓储环境下多目标识别应用场景的问题,提出了一种改进型 Faster R-CNN 目标检测算法.首先,采用 ResNet50 替换 VGG16 作为特征提取网络,以提高模型的检测精度;同时,为兼顾多尺度及小目标物体的检测,引入了特征金字塔网络,形成了残差金字塔特征提取网络 ResFPN;其次,引入了注意力机制,提高输入特征的空间和通道有效信息利用率;最后,使用 RoI Align代替原有的 RoI Pooling,以消除因量化取整而产生的预测框回归误差.在经图像增广处理的自建数据集上进行实验测试,结果表明,提出的改进型 Faster R-CNN 算法在仓储环境下能满足对人员、叉车和托盘的目标检测需求,其平均检测精确度能达到 90.2%.
In order to solve the problem of low accuracy and poor effectiveness of traditional target detection,which is difficult to adapt to the application scenarios of multi-target recognition in warehouse environment,an improved Faster R-CNN target detection algorithm is proposed.Firstly,ResNet50 is used to replace VGG16 as the feature extraction network to improve the detection accuracy of the model.At the same time,in order to take into account the detection of multi-scale and small target objects,a feature pyramid network is introduced to form a residual pyramid feature extraction network called ResFPN.Secondly,attention mechanism is introduced to improve the effective information utilization rate of the input feature space and channels.Finally,ROI Align is used to replace the original ROI Pooling to eliminate the prediction box re-gression error caused by quantization rounding.The experimental tests were conducted on the self-built data set with data augmentation.The experimental results show that the improved Faster R-CNN algorithm proposed in this paper can meet the detection requirements of targets such as people,forklifts and pallets in the warehouse environment with an average de-tection accuracy of 90.2%.
国家科技期刊平台
NSTL
万方数据
