排序方式: 共有3条查询结果,搜索用时 0 毫秒
1
1.
环境车辆是自动驾驶汽车行驶时的主要障碍物之一,对环境车辆的尺寸、位置、朝向等空间信息进行感知对于保障行驶安全具有重要意义。激光雷达点云数据包含了场景中物体表面扫描点的三维坐标,是实现车辆目标检测任务的重要数据来源。结合SECOND与PointPillars方法,提出一种基于体素柱形的三维车辆目标检测模型,利用三维稀疏卷积聚合点云局部特征,在体素特征图上构造柱形并进行特征编码,有效解决点柱形方法缺乏柱形间特征交互问题,增强点云特征的空间语义信息;基于均值池化操作生成锚框点云占据位图并提出一种简单负样本过滤策略,在模型训练阶段筛除预设在无点云覆盖区域的无效锚框,缓解正负样本或难易样本不均衡问题;在目标框回归模块中,使用类别置信度与交并比(IoU)预测值计算混合置信度,改善分类分支与位置回归分支结果不一致问题,并使用KITTI三维目标检测数据集进行模型训练与评估。研究结果表明:在严格判定标准下(IoU阈值设置为0.7),所提算法在简单、中等、困难3种难度级别下分别获得了89.60%、79.17%、77.75%的平均检测精度(AP3D),与SECOND、PointPill... 相似文献
2.
Vehicle positioning is critical for inter-vehicle communication, navigation, vehicle monitoring and tracking. They are regarded as the core technology ensuring safety in everyday-driving. This paper proposes an enhanced vehicle ego-localization method based on streetscape image database. It is most useful in the global positioning system(GPS) blind area. Firstly, a database is built by collecting streetscape images, extracting dominant color feature and detecting speeded up robust feature(SURF) points. Secondly, an image that the vehicle shoots at one point is analyzed to find a matching image in the database by dynamic programming(DP)matching. According to the image similarity, several images with higher probabilities are selected to realize coarse positioning. Finally, different weights are set to the coordinates of the shooting location with the maximum similarity and its 8 neighborhoods according to the number of matching points, and then interpolating calculation is applied to complete accurate positioning. Experimental results show that the accuracy of this study is less than 1.5 m and its running time is about 3.6 s. These are basically in line with the practical need. The described system has an advantage of low cost, high reliability and strong resistance to signal interference, so it has a better practical value as compared with visual odometry(VO) and radio frequency identification(RFID) based approach for vehicle positioning in the case of GPS not working. 相似文献
3.
1