摘要
目标检测技术通过分析传感器采集的图像数据,检测和识别环境中的物体。然而,单一传感器采集的图像在复杂场景中图像特征无法完整表达,造成检测结果出现漏检或误检,导致检测精度下降。但是多传感器采集的图像三维特征冲突又给阈值设定带来干扰。对此,提出一种基于多传感器点云图像数据融合的目标检测算法。通过对多传感器点云数据展开融合处理,提高点云数据质量;采用改进的PointPillars模型将点云图像数据转变为二维图像,解决特征冲突问题。将其输入YOLOv3算法中,通过使用深度卷积神经网络来提取图像特征,并采用多尺度融合的方法快速输出准确的检测结果,提高了检测精度。实验结果表明:所提方法在测试过程中的J指数均保持在0.8以上,在大规模的目标检测过程中表现出较高的检测精度。
Object detection technology detects and recognizes objects in the environment by analyzing image data collected by sensors.However,images collected by a single sensor cannot fully express image features in complex scenes,resulting in missed or false detections in detection results and a decrease in detection accuracy.However,the conflicting three-dimensional features of images collected by multiple sensors also interfere with threshold setting.A target detection algorithm based on multi-sensor point cloud image data fusion is proposed.By fusing multi-sensor point cloud data,the quality of point cloud data can be improved;Using an improved PointPillars model to transform point cloud image data into two-dimensional images,solving the problem of feature conflicts.Input it into the YOLOv3 algorithm,extract image features using deep convolutional neural networks,and use multi-scale fusion methods to quickly output accurate detection results,improving detection accuracy.The experimental results show that the proposed method maintains a J-index above 0.8 during the testing process,demonstrating high detection accuracy in large-scale object detection processes.
作者
文灵
王磊
WEN Ling;WANG Lei(Urumqi Vocational University,Urumqi 830000,China)
出处
《自动化与仪器仪表》
2025年第7期6-9,14,共5页
Automation & Instrumentation
基金
乌鲁木齐职业大学校级课题,WZDSJGG014积极教学法在《安防系统工程技术》课程中的实践与研究
中国职业教育学会2019年规划课题GZYYB2019013新疆建筑设备类专业学生职业技能与职业精神融合研究。
