1 of 29

2 of 29

Evaluating Safety Metrics for Vulnerable Road Users at Urban Traffic Intersections Using High-Density Infrastructure LiDAR System

Prabin Rath, Blake Harrison, Arizona State University

Duo Lu, Rider University, Yezhou Yang, Arizona State University

Jeffrey Wishart, Science Foundation of Arizona, Hongbin Yu, Arizona State University

3 of 29

Traffic Safety with Infrastructure Sensors

  • Over 50% of traffic-related fatalities and injuries occur at intersections, which requires automated traffic safety infrastructure at high-risk zones.
  • Aggressive driving behavior and distracted driving causes many drivers to collide with pedestrians and other vehicles in traffic negotiation.
  • A major goal for safety is to use sensors that can perceive vehicle position and velocities while providing safety to drivers in real-time.

​

Paper # 24SS-0212

3

SAE INTERNATIONAL

4 of 29

Introduction – Using LiDAR Data for Traffic Safety

  • Traffic safety infrastructure has incorporated the use of modern technologies such as LiDAR for real-time vehicle detection.

Paper # 24SS-0212

4

Niraj Altekar, Steven Como, Duo Lu, Jeffrey Wishart, Donald Bruyere, Faisal Saleem, and K. Larry Head. "Infrastructure-based Sensor Data Capture Systems for Measurement of Operational Safety Assessment (OSA) Metrics." SAE International Journal of Advances and Current Practices in Mobility 3, no. 2021-01-0175 (2021): 1933-1944.

SAE INTERNATIONAL

5 of 29

Introduction – Advantages of LiDAR Data

  • With a reported centimeter level error and sampling rate of nearly 1.5 million points per second, LiDAR data have high resolution and accuracy.
  • From point cloud data, objects may be detected using current SOTA deep learning models and tracked using bayesian filters.

Paper # 24SS-0212

5

LIDAR

person

vehicle

SAE INTERNATIONAL

6 of 29

Prior Work – LiDAR-based Safety Metrics Computation

  • From prior work made by Anshuman et al. [1], showed position and velocity measurements from LiDAR may be used to calculate operational safety metrics.
  • Follow up work from Siddarth et al. [2] compared fidelity between infrastructure-based and vehicle-top-based LiDAR setups for metrics computation.
  • Elli et al. [3] used Minimum Distance Safety Envelope (MDSE) safety metrics for vehicle-to-vehicle interactions and compared it with conventional metrics such as Post Encroachment Time (PET).

​

Paper # 24SS-0212

6

SAE INTERNATIONAL

7 of 29

Introduction – Objectives

In this work, we focus on Vulnerable Road Users (VRUs):

  • Pedestrians
  • Bicyclists
  • Micro Mobility Vehicles

Objectives of our work

  • Understand the dynamics of traffic for real-time situational awareness.
  • Predict potentially unsafe situations for VRUs and vehicles.
  • Evaluate real-time safety metrics related to vehicles and VRUs.

​

​

Paper # 24SS-0212

7

SAE INTERNATIONAL

8 of 29

Pedestrian Detection and Tracking

Paper # 24SS-0212

8

Day time pedestrian detection and tracking

SAE INTERNATIONAL

9 of 29

Cyclist Detection and Tracking

Paper # 24SS-0212

9

Night time cyclist detection and tracking

SAE INTERNATIONAL

10 of 29

Micro-Mobility Detection and Tracking

Paper # 24SS-0212

10

Night time micro-mobility vehicle (e-scooter) detection and tracking

SAE INTERNATIONAL

11 of 29

Introduction – Objectives

In this work, we focus on Vulnerable Road Users (VRUs):

  • Pedestrians
  • Bicyclists
  • Micro Mobility Vehicles

Objectives of our work

  • Understand the dynamics of traffic for real-time situational awareness.
  • Predict potentially unsafe situations for VRUs and vehicles.
  • Evaluate real-time safety metrics related to vehicles and VRUs.

​

​

Paper # 24SS-0212

11

SAE INTERNATIONAL

12 of 29

Research Issues

Prior works only consider local and highway intersections with only vehicles.

We extend this work to urban traffic situations for VRU safety analysis.

​

We intend to answer the following key questions

  • Are infrastructure-based LiDARs suitable for analyzing urban traffic scenarios?
  • How can we compute safety metrics for analyzing interactions between vehicles and VRUs?

Paper # 24SS-0212

12

SAE INTERNATIONAL

13 of 29

Data Collection

For our analysis we collect data from an urban traffic intersection at testbed located in downtown Tempe, Arizona

​

​

Paper # 24SS-0212

13

SAE INTERNATIONAL

14 of 29

Data Collection

We collect data during the dusk period, with the objective of capturing the transition from daylight to night.

​

​

​

Paper # 24SS-0212

14

Day time Data Collection

SAE INTERNATIONAL

15 of 29

Data Collection

We collect data during the dusk period, with the objective of capturing the transition from daylight to night.

​

​

​

Paper # 24SS-0212

15

Night time Data Collection

SAE INTERNATIONAL

16 of 29

Data Collection

  • Four batches of data at different time of the day are collected.
  • In total, we collected 18 minutes of traffic data.

Paper # 24SS-0212

16

Data Collection Batches

SAE INTERNATIONAL

17 of 29

Methodology - Detection and Tracking

We use a PV-RCNN [4] model pretrained on KITTI dataset for 3D object detection and Kalman Filters for tracking.

  • High density 128 channel LiDARs capture dense VRU features that are essential for reliable detection.
  • The pointcloud is downsampled using 10 cm voxel grid to ensure uniform point density.
  • Different Kalman Filters are used for vehicles and VRUs as we empirically find the optimal parameters to be highly dependent on object class.

​

Paper # 24SS-0212

17

SAE INTERNATIONAL

18 of 29

Methodology - Detection and Tracking

Paper # 24SS-0212

18

SAE INTERNATIONAL

19 of 29

Object Detection Performance - LiDAR vs Camera

Table below shows the Precision (P) and Recall (R) values for both Camera and LiDAR detections during day and night time. LiDAR shows more reliability.

Paper # 24SS-0212

19

SAE INTERNATIONAL

20 of 29

Methodology - Detection and Tracking

Paper # 24SS-0212

20

(left) Vehicle tracklets and (right) VRU tracklets over a period of 1 hour from our tracking pipeline. Overlaid on a Google Earth image of the Tempe intersection.

SAE INTERNATIONAL

21 of 29

Methodology - Identifying Unsafe Interactions

  • We use the real-time tracking information to identify potentially unsafe interactions between detected objects.
    • We extrapolate the vehicle trajectory using the instantaneous velocity vector times 5 seconds into the future.
    • 2D vector to vector intersections are computed using Bentley-Ottmann algorithm [5] to determine real-time conflict points.
  • All entities involved in such interactions are uniquely identified to be in unsafe situations.

Paper # 24SS-0212

21

SAE INTERNATIONAL

22 of 29

Methodology - Safety Metrics

Paper # 24SS-0212

22

The pink conflict point is the intersection between scaled velocity vectors.

SAE INTERNATIONAL

23 of 29

Methodology - Safety Metrics

  • For every unsafe pair in the intersection, we evaluate real-time safety metrics with respect to the detected conflict point.
    • Post Encroachment Time (PET)
    • Minimum Distance Safety Envelope (MDSE)
  • Time complexity of our approach is reduced compared to exhaustive pairwise checking in prior works.
  • Our approach is faster and more suitable for real-time computation.

Paper # 24SS-0212

23

SAE INTERNATIONAL

24 of 29

Methodology - Safety Metrics

Paper # 24SS-0212

24

vehicle reacting

MDSE

vehicle breaking

Conflict

Point

Predicted Pedestrian Trajectory

MDSE

t2

t1

PET

SAE INTERNATIONAL

25 of 29

Methodology - Safety Metrics

Paper # 24SS-0212

25

VRU-to-vehicle unsafe situation with real-time PET metrics calculation.

SAE INTERNATIONAL

26 of 29

Safety Violations

Paper # 24SS-0212

26

  • A PET violation occurs when the PET value drops below 1.5 seconds.
  • A MDSE violation occurs when the distance to the conflict point drops within the calculated safety envelope.

SAE INTERNATIONAL

27 of 29

Results - Safety Violations

Table below shows the safety metrics violation for VRU-vehicle interactions.

​

Paper # 24SS-0212

27

SAE INTERNATIONAL

28 of 29

Conclusion

Our work underscores the importance of real-time and continuous monitoring systems for VRU safety for urban traffic scenarios.

Our contributions:

  • An LiDAR-based infrastructure monitoring system for detecting potentially unsafe situations.
  • Analysis on the effectiveness of DA metrics for safety of VRUs.
  • Open-source data for the AV community for VRU safety research.

​

Paper # 24SS-0212

28

SAE INTERNATIONAL

29 of 29

References

Paper # 24SS-0212

29

[1] A. Srinivasan, Y. Mahartayasa, V. Jammula, D. Lu, S. Como, J. Wishart, Y. Yang, and H. Yu, “Infrastucture-Based LIDAR Monitoring for Assessing Automated Driving Safety,” SAE Technical Paper 2022-01-0081, 2022.

​

[2] S. Das, P. Rath, D. Lu, T. Smith, J. Wishart, and H. Yu, “Comparison of Infrastructure-and Onboard Vehicle-Based Sensor Systems in Measuring Operational Safety Assessment (OSA) Metrics,” tech. rep., SAE Technical Paper, 2023.

​

[3] M. S. Elli, J. Wishart, S. Como, S. Dhakshinamoorthy, and J. Weast, “Evaluation of operational safety assessment (osa) metrics for automated vehicles in simulation,” tech. rep., SAE Technical Paper, 2021.

​

[4] S. Shi, C. Guo, L. Jiang, Z. Wang, J. Shi, X. Wang, and H. Li, “Pv-rcnn: Point-voxel feature set abstraction for 3d object detection,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 10529–10538, 2020.

​

[5] Bentley and Ottmann, “Algorithms for reporting and counting geometric intersections,” IEEE Transactions on computers, vol. 100, no. 9, pp. 643–647, 1979.

​

​

​

​

SAE INTERNATIONAL