Localization capability demonstration in Dense Urban Secenario During Peak Hours

[liuXinGangChina]

Test condition

1. Autoware Version: 2024.07.17
2. Docker Version: Official Docker 2024.08
3. Launch file: Logging_simulator.launch (only localization and related module been pulled up)
4. Rosbag: Lidar Ouster os1-64-U (mode 1024x10)
5. Traffic status: Rush hour (9.00 am)
6. Traffic Participants: e-bike, bike, motorcycle, bus, truck, car, pedestrian
7. Test route Distance: 5km
8. Duration: 16 minutes
9. Route elements: more than 10 crossroads and zebra crossing, pass through hybrid road which shared by bike and car

Test result

1. Localization pipeline works well under different scenarios, localization failure or crush not detected

2. Path filtered by ekf is smooth and continuous

3. NDT‘s exec time:

   1. Average: 3ms
   2. Max: 14ms
      

4. Rviz screen cast: including localization visualization, dash cam video click here to watch the video
   

5. we also compare the localization pipeline's performance between sunny day and rainy day

Weather	Ndt_calculate_time	video link
Rainy	3-14ms	click here to watch the video
Sunny	3-10 ms	click here to watch the video

[liuXinGangChina]

further comparison between different scenarios

[liuXinGangChina]

We also test the ndt based localization pipeline during peak hour in sunny day， test result is here

So far we have tested localization pipeline in sunny and rainy days during peak hour while the map is built in common days

it turn out to that there are no siginificant difference in calculation time between two weather types

the only difference between those two is during rainy days when ego turn at the road cross，the max exec-time of ndt increase a little from 10 ms to 14ms than sunny day driving

[lntsh854]

Your NDT calculation time is only a few to tens of milliseconds, which is very fast. May I ask what your computing platform is, and what algorithm optimizations have been made compared to the original version of Autoware?

[villaoningfootball]

Did you verify it on a passenger car？

[liuXinGangChina]

Hi @villaoningfootball

The data was collected from our field test vehicle (passenger car) in real scenario ( NanJing of China ).
Most of our test was done using passenger vehicle.
I see you with JD, our team site in NanJing and welcome to contact us.

Happy hacking!

Xingang

[Jinshen-Yan]

The matching speed and accuracy of NDT should mainly depend on the structure of the scene, the number of sampling points and the initial pose. The noise on rainy days may have a relatively small impact on the matching time. Have you verified the stability of ndt when the lidar is severely blocked or there are many dynamic objects?

[liuXinGangChina]

Hi， thanks for your comment @Jinshen-Yan

During the data collection driving, we been through peak hour traffic flow in China NanJing

1. Sometimes ego is surrounf by truck or bus which is much higher than itself, I believe that is what you mean " blocking"
2. In some of the intersections ， you can see there are many dynamic traffic interractors on the road which can be treated as dynamic objects

by the way we use default ndt params of autoware

happy hacking！

心刚