摘要
针对井下照明情况复杂、光线不均匀、背景复杂、行人特征不明显导致基于计算机图形识别的井下行人检测效果不佳这一问题,提出一种基于改进Cascade R-CNN的井下行人检测方法,以Cascade R-CNN为基础,引入Soft-NMS替换传统NMS,充分利用Cascade R-CNN的多阶段检测模型提高检测效果。实验表明:基于改进Cascade R-CNN的井下行人检测方法可有效针对井下特殊复杂情况,在井下行人数据集上获得了91.4%的检测准确率,并使用COCO检测评价矩阵评估模型对改进Cascade R-CNN算法进行了验证,相较于传统Cascade R-CNN算法平均精准度(AP)提升约2%。
Aiming at the problems of downhole lighting,uneven light,single background,and insignificant pedestrian characteristics,which caused poor detection of downhole pedestrians based on computer graphics recognition,a downhole pedestrian detection method based on improved Cascade R-CNN was proposed.Based on R-CNN,Soft-NMS is introduced to replace traditional NMS,and Cascade R-CNN's multi-stage detection model is used to improve the detection effect.Experiments show that the downhole human detection method based on the improved Cascade R-CNN can effectively address the special and complex conditions downhole,and has obtained a detection accuracy of 91.4%on the downhole human data set.The COCO detection evaluation matrix evaluation model is used to improve the Cascade R-CNN.The CNN algorithm has been verified,and the average precision(AP)of the traditional Cascade R-CNN algorithm is improved by about 2%.
作者
袁海娣
YUAN Hai-di(Basic Experiment Teaching Center of Anhui Sanlian College,Hefei 230601,China)
出处
《齐鲁工业大学学报》
2020年第3期68-73,共6页
Journal of Qilu University of Technology
基金
安徽省省级质量工程项目(2018MOOC082)
安徽三联学院院级重点项目(KJZD2019011)
安徽三联学院校级质量工程项目(19zlgc007)。
