基于改进faster RCNN的木材运输车辆检测

Timeber transport vehicle detection based on improved faster RCNN

针对目前森林资源受到盗砍盗伐威胁,相关木材运输车辆行为隐蔽,进而导致无法准确地在交通视频中被识别的问题,提出了一种基于改进fater区域卷积神经网络(faster RCNN)的木材运输车辆检测方法。首先,采用faster RCNN作为基础检测框架,使用金字塔特征网络(FPN)、多尺度训练、锚点框聚类作为基础改进措施;其次,以广义交并比(GIoU)损失函数替换原算法中的smoothL1损失函数作为边界框定位回归的损失函数;最后,计算出在多种实验条件下的模型平均精度均值(mAP),对各种算法进行了对比。实验结果表明,使用GIoU作为损失函数的faster RCNN相比原算法对木材运输车辆检测的平均精度(AP)上升了7.5%,模型平均精度均值(mAP)上升了4.3%;同时,在大型数据集PASCALVOC上,使用GIoU作为损失函数的faster RCNN的mAP达到73.4%,相比其他算法具有明显优势。

Because the problem that the current forest resources are threatened by theft and the related timber transport vehicles’ s activities are hidden,these vehicles can not be accurately identified in the traffic video. A timber transport vehicle detection method based on improved faster Regional Convolutional Neural Network(RCNN)was proposed. Firstly,faster RCNN was used as the basic detection framework,and Pyramid Feature Network(PFN),multi-scale training,anchor frame clustering were adopted as the basic improvement measures. Secondly,the smoothL1 loss in the original algorithm was replaced with Generalized Intersection over Union(GIoU)loss function as the loss function of the boundary box regression.Finally,the mean Average Precisions(mAPs)of the model under various experimental conditions were calculated,and the various algorithms are compared. The experimental results show that the Average Precision(AP)of the faster RCNN with GIoU as the loss function is 7. 5% higher than that of the original algorithm,and the model mAP is increased by 4. 3%. At the same time,a comparative experiment was designed on PASCALVOC dataset. The faster RCNN with GIoU loss function has a mAP value of 73. 4%,which has significant advantages over the comparison algorithms.