Factors associated with pedestrian-vehicle collision hotspots involving seniors and children: a deep learning analysis of street-level images

ABSTRACT

This study aimed to examine the factors associated with pedestrian–vehicle collision hotspots involving seniors and children. For the empirical analysis, we first quantified street-level images of collision hotspots involving seniors and children and non-collision hotspots in the Seoul Metropolitan Area, Korea, using deep learning analysis. Thereafter, we examined the risk factors associated with collision hotspots through logistic analyses. This study has two major findings. First, the effects of risk factors (e.g. share of sky and green space) differ between collision hotspots involving seniors and children. Second, some pedestrian safety treatments (i.e. traffic lights and sidewalks) are positively associated with collision risks. The findings suggest that varied approaches to enhancing pedestrian safety among different age groups should be considered for more effective pedestrian safety interventions. In addition, the quality of pedestrian safety measures should be examined to improve pedestrian safety for seniors and children.

KEYWORDS:

• Pedestrian
• collision
• senior
• child
• deep learning
• street-level image

Disclosure statement

No potential conflict of interest was reported by the author(s).

Notes

1 We did not use the 2020 and 2021 data because the pedestrian–vehicle collision environment might be abnormal owing to the COVID-19 pandemic that began in 2020 and lasted until 2021.

2 In the case that there was no GSV image at the randomly selected coordinate for a non-collision hotspot, we moved to the next randomly selected coordinate to collect GSV images.

3 While we used yearly data of the collision hotspots between 2013 and 2019, the GSV images in the collision hotspots were not available every year. Given this limitation, we collected the GSV images that minimized the time difference between the year of a collision hotspot and the year of its associated GSV images taken. For example, when GSV images were available for years of 2015 and 2018 in a collision hotspot in 2019, we collected the GSV images that were taken in 2018.

4 To estimate the performance of the SS model in this study, we evaluated the Intersection of Union (IoU) value on the ADE20K validation dataset. The IoU is the area of overlap between the predicted segmentation and the ground truth divided by the area of union between the predicted segmentation and the ground truth. It is a commonly used evaluation metric of SS performance. Appendix 1 shows the mean IoU (mIoU) performance of PSPNet, our SS model, which shows much higher performance for the sky, green space, road, and sidewalk classes than for the other classes in the ADE20K dataset. This demonstrates that the SS model can effectively extract the chosen variables from a scene.

5 Given the collision data do not include pedestrian–bicycle collisions, we did not include bicycles when counting the number of vehicles.

6 The YOLOv5 model can count persons shown in an image that are not specifically pedestrians. In other words, the OD method counts those who are riding bicycles or motorcycles as persons but cannot separate pedestrians among persons. Thus, we addressed this issue by utilizing the IoU values generated by the detection model. IoU connotes the degree of overlapping between two bounding boxes of objects. The IoU value is calculated by dividing the area of intersection between the two boxes by the area of their union. Consequently, the resulting value ranges from 0 to 1, with a higher value indicating a large match between two bounding boxes. We measured all IoU values between persons and motorcycles or bicycles. If the IOU value exceeded the predefined threshold θ, we excluded that person, thereby sorting out pedestrians from all persons. A higher IoU value between a person and a motorcycle or bicycle indicated that the person was likely a motorcyclist or bicyclist rather than a pedestrian. We set the IoU threshold θ as 0.4 empirically, which yielded satisfactory results in qualitative evaluation.

7 To check the validity of our object detection model, we evaluated the performance for our target classes with the COCO 2017 validation dataset (Lin et al., 2014). We used the mean average precision (mAP) metric as in the Pascal VOC 2007 challenge (Everingham, Van Gool, Williams, Winn, & Zisserman, 2010) for the quantitative evaluation of the object detection model. Appendix 2 shows the performance of the YOLOv5 object detection model, indicating that the object detection model can easily detect traffic-related objects, especially persons and vehicles.

Additional information

Funding

This paper was supported by AI Convergence Research Fund, Sungkyunkwan University, 2021.