To train the computer of a car to drive itself, images from a camera attached to the car and its corresponding steering and other outputs can be feed to a convolutional neural network to learn its own outputs when it encounters similar data. This technique is called behavioral cloning.
The goals / steps of this project are the following:
This is project #3 of my Self Driving Car Projects
The environment can be created with CarND Term1 Starter Kit
The Udacity Unity Car Simulator can be downloaded here. Ignore the instructions to install Unity, as Unity loads automatically with executable. For Linux, you may need to
chmod 777 linux_sim.x86_64 it if they still have not done that already.
The project includes the following source files:
train.py- train the convolution neural network model and save it
datapaths.py- defines image and csv file paths
pd_util.py- helper functions
drive.py- Udacity provided script for driving the car in autonomous mode
video.py- Udacity provided script for generating video from recorded images
The model is trained by a dataset of images from the car camera’s point of view of the road and its corresponding steering angles. These are specified in
datapaths.py. The recorded dataset is stored in a separate repo.
You can train your model with above recorded dataset, use Udacity’s recorded dataset, or generate your own data via the record feature in the simulator. To do so, first start the simulator.
640x480 for “Graphics” and
Fastest for “Quality”.
Select “Training Mode” and click the record button when you like to start recording. It will ask you where the video images will be saved to. You will need to modify the paths in
datapaths.py to where the images and csv is saved to.
Once you have recorded at least one lap of images, you can start the training via:
This will save the model to
model.h5 when complete. Training takes a few hours depending on the GPU you use.
When training is complete, you can see it in action by selecting “Autonomous Mode” in the simulator and run the model via:
python drive.py model.h5
To produce a video recording of the run from the car camera’s point of view, hit the record button in the simulator and stop it when you like to stop. Then run:
python video.py RECORDED_IMAGES_PATH
This will create a mp4 video output of car’s point of view of the run.
The model was trained and validated on different data sets to ensure that the model was not overfitting. It uses an adam optimizer and includes ELU (Exponential Linear Unit) layers to introduce nonlinearity. The trained model is able to drive autonomously around the track without leaving the road after some modification to
drive.py (more on this in Results section below).
A combination of center lane driving (1 lap), recovering from the left and right sides of the road (~1 lap), and going slowly on sharp turns (1 lap).
The training data is split into a training and validation set. This helps to gauge how well the model is working and to determine if the model was over or under fitting.
The final architecture produced the following losses:
+ Epoch 1: .0326 loss; .0080 validation loss + Epoch 2: .0081 loss; .0074 validation loss + Epoch 3: .0148 loss; .0079 validation loss
Epoch 3 seems to be getting worst so it indicates possible overfitting.
The final model is a run with 2 epochs. Although running it with 3 epochs seems no difference.
The final step was to run the simulator to see how well the car was driving around track one. There were a few spots where the vehicle fell off the track. To improve the driving behavior in these cases, I tried gathering better training data and also tried training with the Vgg16 architecture. Since training takes hours even for just 2 - 3 epochs, it is difficult to experiment with too many different parameters and layers.
The final model architecture consists of the Nvidia convolution neural network with the following layers and layer sizes:
To capture good driving behavior, one lap on track 1 using center lane driving under medium to medium-high speed is recorded.
I then run another lap running full speed on straight roads, but slow down on tough turns so that more images and data points can be captured on the trouble turns.
These tow laps make up about 6000 records.
I then recorded the vehicle recovering from the left side and right sides of the road back to center so that the vehicle would learn to recover when it is heading off road for both straight roads and sharp turns.
These make up about 3000 records.
These are combined together giving a total of over 9200 records.
The first step is to remove those of speed less than 1 mph.
This gives a total record count of 9075.
Next the number of records are doubled by copying the records but indicating that the center image will receive a mirror operation later in the pipeline and changing its steering value to its negation.
This gives a total record count of 18,150.
Next the data is randomly shuffled.
These steps are done in the middle of
log = pdu.filter_gte(log, 'speed', 1) mlog = pdu.mirror(log, 'center', 'steering') mlog = pdu.shuffle(mlog)
This data set is split into 80% training and 20% validation.
This gives a training data set of 14520 records, and 3630 in validation.
The training set and validation set is generated with a batch generation function (
train.py generates this function). An image processor function (
np_util.py generates this function) is passed to it that is defined to do the following processing steps:
Since the images are 320 x 160, the resulting image is 300 x 64.
Resize to 64 x 64. Since the height is already 64 pixels, only scaling is in the horizontal direction.
Convert to BGR and normalize the pixels.
Flip the image if is was marked to be flipped above.
Adding more epochs to training did not seem to help getting the car to steer correctly at sharp turns reliably. Usually within a few laps, it will steer into a sharp turn the wrong way (too straight) and can not recover. After not getting any better results from completely different training recordings and more tough turns recordings, and trying different parameters and layers, I had to think a bit outside the box. Instead of getting a better model to make these turns better, what if slow down and not run the simulation at full throttle?
To do so, I modify
drive.py with the following throttle settings:
Here is a summary of what throttle settings work and its failure rates:
|throttle||max speed||went off course in|
|0.3+||30 mph||1 or 2 laps|
|0.2||23 mph||3 laps|
|0.15||17 mph||5 laps|
|0.1||11 mph||Did not go off course in over 1.5 hrs, ~50 laps|
So throttle of 0.1 is good. But thought I can do better.
What if I amplify the model by upping the speed at straight roads and slowing down and steering more than model suggests at sharp turns?
I set the throttle at 0.2 normally and when the absolute value of steering input is over 4, I brake at .0001 x absolute steering value. This slows it down to ~11 mph at sharp turns. The steering is multiplied by 1.5 if absolute steering input is over 4 and multiplied by 1.8 if over 6.
This setting runs the laps at a much higher speed and it ran over 1.5 hrs, or over 66 laps without going off course.
Sorry, I forgot to upload the video recording before I deleted it. It is not much different than many others you can find on youtube, so just search there if you like to see how the simulated result looks like.
I have a love and hate relationship with this project. It is fun to see the model run forever without going off course at the end. But with a slow GPU, the feedback process is poor. I want to see the model work on the very difficult track two. To do so, I will need to either invest in new GPU or use cloud offerings to continue trying different the model architecture, layers, and parameters.
Next up is project 4 - Advanced Lane Line Detection.