Pedestrian detection method in rail transit scenes based on fusion of 3D point clouds and images
Safety incidents caused by pedestrians illegally intruding onto railway tracks occur frequently in rail transit scenes, significantly affecting the safe operation of trains. Utilizing single sensor data for pedestrian detection often leads to low recall rates of detection results and lack of categor...
Saved in:
| Main Author: | |
|---|---|
| Format: | Article |
| Language: | zho |
| Published: |
Editorial Department of Electric Drive for Locomotives
2024-05-01
|
| Series: | 机车电传动 |
| Subjects: | |
| Online Access: | http://edl.csrzic.com/thesisDetails#10.13890/j.issn.1000-128X.2024.03.105 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Safety incidents caused by pedestrians illegally intruding onto railway tracks occur frequently in rail transit scenes, significantly affecting the safe operation of trains. Utilizing single sensor data for pedestrian detection often leads to low recall rates of detection results and lack of category or orientation information in the results, which cannot meet practical field requirements. To address these issues, this paper proposed a pedestrian detection method based on the fusion of 3D point clouds and images in rail transit scenes. This method first employed a deep learning model trained on a constructed dataset of pedestrian data in rail transit scenes to detect pedestrians separately in 3D point clouds and images. Subsequently, based on the principle of spatial position consistency of the targets in 3D point clouds and images, the rotation and translation matrix between the LiDAR and camera was solved. Finally, the 3D point cloud object detection results were projected onto the image coordinate system. To solve the issues of misalignment between multiple adjacent targets and the one-to-many relationship between detection results, the intersection over union ratio and center point distance between the two detection results were calculated as fusion constraints, enabling more accurate pedestrian detection. Experimental results using data acquired from the field demonstrate that, compared to detection results from separate data of 3D point clouds and images, while maintaining timeliness, this method improves the recall rate by 4.5% and 5.5%, respectively, effectively reducing the risk of safety accidents caused by missed pedestrian detections, meeting the demand for pedestrian detection during actual train operations. |
|---|---|
| ISSN: | 1000-128X |