In the test-field calibration,multi-azimuth stereo image pairs areproduced of the outdoor large control-field by the stereo-vision system under cali-bration.While in the analytical processing,the relationship between ...In the test-field calibration,multi-azimuth stereo image pairs areproduced of the outdoor large control-field by the stereo-vision system under cali-bration.While in the analytical processing,the relationship between image pairsis adopted as a constraint condition,which ensures the stability and quality of thecalibration results.This paper introduces the deduction process of the constraintconditions.展开更多
为了提升汽车辅助系统对前方障碍物的检测效果并进一步获取精确的距离信息,文章提出了一种基于改进YOLOv8s的交通场景障碍物检测与双目测距方法。该方法以YOLOv8s(You Only Look Once v8s)网络为基础,首先在Backbone中引入EMA注意力机制...为了提升汽车辅助系统对前方障碍物的检测效果并进一步获取精确的距离信息,文章提出了一种基于改进YOLOv8s的交通场景障碍物检测与双目测距方法。该方法以YOLOv8s(You Only Look Once v8s)网络为基础,首先在Backbone中引入EMA注意力机制,以提高目标检测精度;其次将Neck中的PANFPN网络替换为ASF(Attentional Scale Sequence Fusion)网络,并采用DIoU优化损失函数;在特征匹配算法ORB的基础上,利用RANSAC算法剔除误匹配的点对。通过在KITTI数据集和实际交通场景中的实验,结果表明,在20 m的距离范围内,改进后的YOLOv8s网络对汽车、行人和非机动车3类障碍物的检测mAP(mean average precision)达到了91.1%,提高了4.8%,同时测距的平均误差仅为1.55%。展开更多
文摘In the test-field calibration,multi-azimuth stereo image pairs areproduced of the outdoor large control-field by the stereo-vision system under cali-bration.While in the analytical processing,the relationship between image pairsis adopted as a constraint condition,which ensures the stability and quality of thecalibration results.This paper introduces the deduction process of the constraintconditions.
文摘为了提升汽车辅助系统对前方障碍物的检测效果并进一步获取精确的距离信息,文章提出了一种基于改进YOLOv8s的交通场景障碍物检测与双目测距方法。该方法以YOLOv8s(You Only Look Once v8s)网络为基础,首先在Backbone中引入EMA注意力机制,以提高目标检测精度;其次将Neck中的PANFPN网络替换为ASF(Attentional Scale Sequence Fusion)网络,并采用DIoU优化损失函数;在特征匹配算法ORB的基础上,利用RANSAC算法剔除误匹配的点对。通过在KITTI数据集和实际交通场景中的实验,结果表明,在20 m的距离范围内,改进后的YOLOv8s网络对汽车、行人和非机动车3类障碍物的检测mAP(mean average precision)达到了91.1%,提高了4.8%,同时测距的平均误差仅为1.55%。
