More Accurate Representation of Pedestrians in Point Cloud Data

[GoodarzMehr]

Is your feature request related to a problem? Please describe.
The shape of pedestrians in point cloud data is based on the collision of lidar rays with the capsule components of the Physics Asset. This can be very inaccurate and could be an issue when training a model on CARLA point clouds and evaluating it on real-world data.

Describe the solution you'd like
It would be great to add a setting to CARLA where the user can toggle the collision calculations between the Physics Asset capsule components and the pedestrian's trimesh. This obviously comes with a performance penalty but the user will have a choice to use the Physics Assets for faster calculations or the trimesh for slower but more accurate calculations.

Describe alternatives you've considered
For my case, I've turned on Enable Per Poly Collision in both the mesh component of BP_Walker and for each skeletal mesh. It does slow down the simulation, especially when there are a lot of pedestrians in the scene (on a 4090 the total time for all pedestrian computations (not the time added because of using the trimesh instead of the Physics Asset, don't have an estimate for that) becomes ~ $0.001n ^ 2 + 0.2n$ ms where $n$ is the number of pedestrians in the scene.

Additional context
Point cloud data when using the trimesh for calculations: