The problem of perception direction to predict the velocity of objects

[hzkkk1997]

When using the fusion method of LiDAR and Radar to conduct autoware real vehicle test, we found that for stationary objects, when the vehicle accelerates or decelerates, there will be a predicted speed error of about 3km/h. How can we further optimize this problem?
ps: This is a related question(https://github.com/orgs/autowarefoundation/discussions/5321#discussion-7298001) from my partner, This seems to affect decision-making and planning when turning. How does autoware deal with it? thanks a lot!

[scepter914]

I guess your problem separate into two problem.

• Latency

The output from the radars has latency especially for velocity estimation.
It is difficult to handle because we don't know how long the latency of radars, which is often defined by devices specification, but one possible solution is that you use delay compensation in tracking layer.

https://github.com/autowarefoundation/autoware.universe/blob/8e0d40b8510ff69c4653842433dd4d4b284beb79/perception/autoware_multi_object_tracker/config/multi_object_tracker_node.param.yaml#L15

• The error of velocity compensation

The node of radar_tracks_msgs_converter estimate absolute velocity of objects by using radar doppler velocity and ego velocity.
The velocity estimation has the error and I think it stands out when there are stationary objects when the vehicle start and stop with latency problem as I said.

For this, I give two other possible solutions.
First, I recommend that you should use autoware_object_velocity_splitter if you use LiDAR-Radar fusion.
When you use the objects of high velocity from the output and delete the objects of low velocity, you can filter low velocity which cause noise for LiDAR-Radar fusion.

Second, I recommend if you use directory radar objects for detection output, then you make new node such as static_object_validator.
It is a node to delete small velocity and change from small velocity into 0 velocity.

I think these solution is up to your system, so I suggest you consider about these solution and choose one or adapt one of these for your system.

[hzkkk1997]

@scepter914 Thank you for your reply, I will study it.