Academia.eduAcademia.edu

Figure 5 – uploaded by Hongbo Zhang

See full PDFdownloadDownload figure
of view is necessary. Simply increasing the upper field of view is not an optimal strategy. In no surprise, th increase of upper field of view is associated with more visibility of buildings shown in Figure 7. This maybe useful for visual tasks involving building detection and classification, but proven not extremely helpful for the intersection classification. In comparison to the upper field of view, the field of view visualization suggests tha‘ the 5 to 15 degree of lower field of view corresponds to the missing of large details of the point cloud data a: shown in Figure 8. In a surprise, the lack of such details does not lead to the deterioration of the classificatior performance, which is likely caused by the fact that the missing details do not significantly impact the important features useful for intersection classification.

Figure 7 of view is necessary. Simply increasing the upper field of view is not an optimal strategy. In no surprise, th increase of upper field of view is associated with more visibility of buildings shown in Figure 7. This maybe useful for visual tasks involving building detection and classification, but proven not extremely helpful for the intersection classification. In comparison to the upper field of view, the field of view visualization suggests tha‘ the 5 to 15 degree of lower field of view corresponds to the missing of large details of the point cloud data a: shown in Figure 8. In a surprise, the lack of such details does not lead to the deterioration of the classificatior performance, which is likely caused by the fact that the missing details do not significantly impact the important features useful for intersection classification.

Related Figures (10)

Table 1: LiDAR design parameters. The training of Lidar data is subjected to the general pipeline of the PointNet++.°! More specifically, for each condition, e.g. 10 degree of upper field of view or 100 meters detection range, 1000 self-driving trips were simulated. The corresponding point cloud data for these trips are used as the input data. The data is further splitted into 60 % as training and 40 % as testing. 200 epochs are used for training the point cloud data for classification of intersection versus non intersection. Adam optimizer is used for optimization of the PointNet++ learning process. 1024 number of points are used per batch for the training process. Early stop is used for prevention of overfitting.
Table 1: LiDAR design parameters. The training of Lidar data is subjected to the general pipeline of the PointNet++.°! More specifically, for each condition, e.g. 10 degree of upper field of view or 100 meters detection range, 1000 self-driving trips were simulated. The corresponding point cloud data for these trips are used as the input data. The data is further splitted into 60 % as training and 40 % as testing. 200 epochs are used for training the point cloud data for classification of intersection versus non intersection. Adam optimizer is used for optimization of the PointNet++ learning process. 1024 number of points are used per batch for the training process. Early stop is used for prevention of overfitting.
Figure 1: The overall architecture of the end-to-end lidar simulation. The process starts from adjusting different parameters of the LiDAR and obtaining point cloud data from CARLA simulator. The point cloud data are then subjected to classification using PointNet++.
Figure 1: The overall architecture of the end-to-end lidar simulation. The process starts from adjusting different parameters of the LiDAR and obtaining point cloud data from CARLA simulator. The point cloud data are then subjected to classification using PointNet++.
Figure 2: The classification accuracy with different number of points of point cloud.
Figure 2: The classification accuracy with different number of points of point cloud.
the results of classification suggest that the increase of such detailed features does not help to increase the overal accuracy of the classification task.
the results of classification suggest that the increase of such detailed features does not help to increase the overal accuracy of the classification task.
Figure 4: The classification accuracy with different LiDAR detection range.
Figure 4: The classification accuracy with different LiDAR detection range.
Figure 6: The classification accuracy with different vertical upper field of view angles.
Figure 6: The classification accuracy with different vertical upper field of view angles.
Related topics:
Lidar ClassificationSelf DrivingSelf-Driving Cars
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved