Resolution-Aware Roadside LiDAR Perception and A Real-World Testbed
Arizona State University
Dajiang Suo, Assistant Professor
1
10/09/2025
Background and Motivation
2
Chattanooga City Roadside LiDAR deployment over 120 intersections
Image source: Ouster
Roadside LiDAR studies from Academia
[Li 2022]
[Xu 2020]
Why Resolution Matters�- Tradeoff between Cost and Resolution
3
Cost
Camera
2D Resolution
Performance
3D Resolution
From 2D to 3D (Spatial) Resolution
4
Traffic camera
Roadside LiDAR
Credits to SourceSensors
Credits to Agarwal
Real objects
2D representation
3D representation
Why Resolution Matters�- A Deep Learning Perspective
5
MLP
Preprocessing
Representation & Partitioning
CNN
Feature encoding
Backbone
MLP
Task heads
Results
From 2D to 3D (Spatial) Resolution
6
Traffic camera
Roadside LiDAR
Credits to SourceSensors
Research question: To what extent do sensors that provide 3D spatial resolution (e.g., LiDAR) complement and/or outperform 2D cameras in various roadside sensing tasks?
Resolution Impact on Downstream ITS Tasks
7
Credits to Goget.com
Image source: MAG & ADOT
[Toe 2024]
A Real-World Testbed for Studying the Impact of (3D) Spatial Resolution of LiDAR
8
Riggs Rd and S. Dobson Rd at Sun Lakes, AZ
(128/64/16 channels)
Traffic Camera View of the Intersection Testbed
9
(LiDAR) Point Cloud Projection on (Camera) 2D Images
10
16 Channel
128 Channel
64 Channel
LiDAR Resolution and Point Cloud Distribution
11
128 Channel
64 Channel
*He, Y., Cao, P., Suo, D. and Liu, X., 2024. A joint optimization of beam distribution and deployment for roadside LiDAR systems to maximize vehicle perception. IEEE Transactions on Intelligent Vehicles.
LiDAR Resolution and Point Cloud Distribution
12
*He, Y., Cao, P., Suo, D. and Liu, X., 2024. A joint optimization of beam distribution and deployment for roadside LiDAR systems to maximize vehicle perception. IEEE Transactions on Intelligent Vehicles.
What we have on site
LiDAR (Top) View of the Intersection Testbed
13
16 Channel
128 Channel
64 Channel
Impact Analysis on Downstream Tasks
14
[hitech 2024]
Task I. Object Detection (preliminary results)
15
Car
Van
Motorcycle
Trailer
Bus
Truck
Camera Under Off-Nominal Scenarios
16
Glare Scene
Night
Trucks mistakenly identified as cars
Due to night lighting issues, a non-existent truck was identified
Preliminary Results from On-site Experiments
17
| (Average) Detection Accuracy for Various Traffic Participants | ||
Distance to Intersection Center | Cameras | Mid - Resolution LiDAR | High -Resolution LiDAR |
0-50m | 90.8% | ↑7.8% | ↑8.1% |
50-70m | 76.1% | ↑17.6% | ↑20.9% |
>70m | 56.7% | ↓40.7% | ↓36.2% |
Preliminary Results from On-site Experiments
18
| (Average) Detection Accuracy for Various Traffic Participants | ||
Distance to Intersection Center | Cameras | Mid - Resolution LiDAR | High -Resolution LiDAR |
0-50m | 90.8% | ↑7.8% | ↑8.1% |
50-70m | 76.1% | ↑17.6% | ↑20.9% |
>70m | 56.7% | ↓40.7% | ↓36.2% |
Comparison Between Hi- and Med-res LiDARs
19
The figure shows the detection accuracy of 128-C LiDAR and 64-C LiDAR for different types of objects in the range of 0-70m.
128-C
64-C
Motorcycle
Expected Outcomes
20
Open Data Platform for multi-resolution sensing
01
Facilitate AI Algorithms Development
03
Best Practices & Lessons Learned
01
02
Benchmark for Roadside Sensing
03
04
Q&A
21