Hey Steemians , this is another video in the MIT Race Car Series project.In this video we used a trained Auto-pilot model.The model was trained on a track such that it would follow the dotted dashed line in the middle of the track.
In the following video we ran the autopilot to see how well it reacts when certain obstacles are present in it's path.
Our Initial findings shows us that the autopilot has learn't how to distinguish other car's that are in the path,additionally the autopilot has also learnt how to reverse in order to avoid an obstacle if one is present.
In order to train the car you need to run car.py (train) and then drive a few laps.When the car is driving it is essentially capturing images on a frame by frame basis .This was important for our project as it tells us how much training data we have.
Once we have collected enough training data we can run the training algorithm which will run for approximately 50 or so iterations.The training model that we used was able to distinguish the different images on the basis of Pixels in the image.The training time on average is approximately 40 minutes or so.
Some Key points we learnt
| Images Collected | Training Time | Iterations Completed | Model Accuracy |
|---|---|---|---|
| 20000 | 39 min | 25/100 | 25% |
| 40000 | 44 min | 43/100 | 45% |
| 50000 | 55 min | 65/100 | 79% |
On an average about 50000 images produced the best results.and therefore based on our leanings I would recommend collecting at-least 50000 images to get the best results.
Note : The technique used for Training a model is called perceptron algorithm,i'll do another post to show you guys how it works.
Here you can gain access to the code
Below is how you would run the training scripts if you were to do it on you're own.
Usage:
car.py (drive) [--model=]
car.py (train) (--tub=) (--model=)
car.py (calibrate)
