Yes, For sample I've done for few instances using drivingScenarioDesigner tool.

Hi @robertogl , Could you please help me on proceeding with further objective

Hi @DHEEPAK29

Happy new year! Sorry for the late response, but I am finally back.

I am not sure how many instances you have created but you would need enough to be able to train a NN model.
Then you should use data from the simulated sensors to understand whether the behavior of external objects, like pedestrian or other vehicles, is safe or not so that you can label the data (for example from a simulated camera images, or lidar, or both) and start the training. To characterize the objects' behavior as safe or not safe you will need to have info about their motion, so output from a tracking algorithm might be needed.

After you have trained the model you will have to test it in a new test scenario and and check its performances.

I hope this gives you some inputs to move your project forward. Please let me know if you need anything else.

Hi @robertogl, Sorry, I was in a rush for few weeks. As you told I've created few scenarios, further I'll elaborate multiple scenario. For labelling the Features, We have 2 either safe or unsafe. So, If I have many obstacles in a single scenario for example a pedestrians or a barrier. Both of them are different entity, For Pedestrian, The vehicle should stop till the person passes and for barrier the vehicle can take a turn if optimum space is available in the road, Thus, Two actions are entirely different. So, Is it sufficient to classify the instances with only 2 labels i.e., Safe & Unsafe. Can you reflect your idea on this.

image

Hi @DHEEPAK29,

That is a very good point. I think using just the 2 labels, 'Safe' and 'Unsafe' it is a great start and they are good enough to decide whether continue via the planned trajectory/task or change maneuver, respectively. In case of Unsafe state, there might be different reactions as you correctly pointed out, but that can be decided based on the obstacle's type and/or motion. However, I think that is normally taken care of by a different layer, which includes object detection. Does it make sense?

Hi @VoidXia,

I totally understand your concern. I am not aware of any dataset out there but that does not mean that some may not exist already.
Regarding creating a new one from a simulation environment, we were thinking that that process could be, at least partially, automated, by changing the environment by adding or removing objects, changing motion, behavior, etc.

We have many tracking algorithms in our Sensor Fusion and Tracking Toolbox. A simple Kalman Filter could work for objects with linear motion models but we have more advanced multi-object trackers and many examples for autonomous systems.
Perhaps you can start from one of these examples and adapt it to your application.

Yeah, I agree Roberto. You can track using the Sensor Fusion and Tracking Toolbox and the links that he provided are where I would start.

Got it, thanks. You can paste all that information in the README file. If you are familiar with markdown you can have a similar nice formatting. Also, no need to zip the files for GitHub.

Oh I forgot...There is no license in your repository. Would you be willing to release your code under an open source license (2-clause BSD or MIT)?

Yes i am ok to release it as open source license with MIT and i will upload all the info in readme file. Do you need any other help for evaluating my solution.

@robertogl Is my github repo ok for now?

I see that you currently have the 'Apache-2.0' license, can you change it to 2-clause BSD or MIT?
Also, it would be great if you could also add pics and videos in the README file.
Thanks!