摘要
车辆乘员数量检测研究对推进HOV车道建设具有重要意义,本文以Faster RCNN网络模型为基础,结合多光谱红外系统获得的汽车驾驶室图像来展开研究。现有数据中因过曝、曝光不足等问题导致图像中目标特征差异大,网络检测的准确率不高,对此采用变形结构的卷积计算提高特征单元的感受野和目标边缘信息的表述能力,采用变形的ROI-Pooling来增强ROI区域特征映射后的特征表述,增强网络的泛化能力。针对在多乘员情况下,因乘员间遮挡导致的错检、漏检等问题引入KL损失,同时采用Soft-NMS与方差投票相结合的方式来改善NMS滤除重复目标框过程的合理性,提高了位置回归的合理性和重叠目标的预测能力,整体检测的准确率得到提高。实验结果表明,本文网络在不同数量下算法检测的准确率得到提高,基本可以满足行业规定大于80%的要求。
The research on vehicle occupant number detection is of great significance to promote the construction of HOV lanes.Based on the Faster RCNN network model,this paper combines the multi-spectral infrared system to obtain the image of the car cab.In the existing data,due to overexposure and underexposure problems,the target feature differences in the image are large,and the accuracy of network detection is not high.For this,the convolution calculation of the deformed structure is used to improve the ability to express the receptive field of the feature unit and the target edge information.Use deformed ROI-Pooling to enhance the feature expression after ROI feature mapping and enhance the generalization ability of the network.In the case of multiple occupants,the KL loss is introduced due to misdetection and missed detection caused by occlusion among the occupants.At the same time,the combination of Soft-NMS and variance voting is used to improve the rationality of the NMS filtering process of repeated target frames and improve in addition to the rationality of position regression and the prediction ability of overlapping targets,the overall detection accuracy has been improved.The experimental results show that the accuracy of the algorithm detection of the network under different numbers in this paper has been improved,and can basically meet the requirements of the industry regulations greater than 80%.
作者
金鑫
胡英
JIN Xin;HU Ying(College of Marine Electrical Engineering,Dalian Maritime University,Dalian 116026,China)
出处
《激光与红外》
CAS
CSCD
北大核心
2021年第1期52-58,共7页
Laser & Infrared
基金
国家自然科学基金项目(No.61973049)资助。