Using Deep Learning to Clone Driving Behavior


The goal of this project was to utilize deep neural networks and convolutional neural networks in order to clone driving behavior. Udacity provided a simulator application where we can steer a car around a track for data collection. The collected data is comprised of image data captured from a car’s set of cameras as well as steering angles. This is used to train a neural network and then use the model to drive the car autonomously around the track — also using the simulator.


The final content of the project is available in my personal repository in Github.

Collecting training data

The dataset specified in the assignment is composed by:

TheIMGdirectory containing frames from cameras positioned at the center, left and right (front) of the car. Each file is named accordingly.

A CSV file (driving.log) where each entry contains:

Image for post
Image for post
Sample excerpt from driving.log

Udacity has provided a training dataset that included 8036 entries, roughly 25,000 images for all the three cameras. The simulator application also doubles as training tool that allows the capture of driving data in the format described above. In order to generate the data you’re supposed to drive the car around the track using the keyboard controls, akin to typical computer games.

This data capture method using the keyboard isn’t ideal as it takes a lot of gaming expertise and fine control to make the ride smooth. I had a hard time capturing data so I decided to adapt an XBox One controller to do the same job. I installed an open source driver that allows Xbox One controllers to be connected to a macOS laptop using a micro USB cable.

Image for post
Image for post

I then proceeded to capture my own data by driving the car around Track 1 a few laps and then doing the same in the opposite direction. I also recorded data recovering the car from strange situations, like being on the edge of the road, back to the center of the lane.

Model architecture


The implementation of the convolutional neural network model for this project use the Keras which is a high-level API, that runs on top of TensorFlow.


LeNet is already known to us as the convolutional neural network architecture used in the German Traffic Sign Classifier project. It seemed like a good starting point as we were already familiar with it. Instead of 43 classes (or 10 from the MNIST examples) we had to adapt it to input the camera images and output the steering value.

Image for post
Image for post
Classic LeNet Architecture


The third model to be tried was the one based on a paper published by the research team at NVIDIA. This is the network as described by the team.

The first layer of the network performs image normalization. The normalizer is hard-coded and is not adjusted in the learning process. Performing normalization in the network allows the normalization scheme to be altered with the network architecture, and to be accelerated via GPU processing.

The convolutional layers are designed to perform feature extraction, and are chosen empirically through a series of experiments that vary layer configurations. We then use strided convolutions in the first three convolutional layers with a 2×2 stride and a 5×5 kernel, and a non-strided convolution with a 3×3 kernel size in the final two convolutional layers.

We follow the five convolutional layers with three fully connected layers, leading to a final output control value which is the inverse-turning-radius. The fully connected layers are designed to function as a controller for steering, but we noted that by training the system end-to-end, it is not possible to make a clean break between which parts of the network function primarily as feature extractor, and which serve as controller.

And this is the summary of the network as displayed by the Keras API.

Image for post
Image for post
NVIDIA network architecture as implemented

The difference from the paper original model is that I used images of (64,64,3)after resizing the input files. We’ll get into more details in the Dataset Augmentation section.

Note on overfitting: I tried adding dropout layers in the network to reduce overfitting, but it had negative impact in the track tests, with the car being unable to finish a full lap. The layers were added after each fully connected layer in the model.

def model_NVIDIA():
model = Sequential()
model.add(Lambda(lambda x: x/127.5 - 1., input_shape=(64, 64, 3)))
model.add(Convolution2D(24, 5, 5, subsample=(2,2), activation='relu'))
model.add(Convolution2D(36, 5,5, subsample=(2,2), activation='relu'))
model.add(Convolution2D(48, 5, 5, subsample=(2,2), activation='relu'))
model.add(Convolution2D(64, 3, 3, activation='relu'))
model.add(Convolution2D(64, 3, 3, activation='relu'))
model.add(Dense(1164, activation='relu', name='FC1'))
model.add(Dense(100, activation='relu', name='FC2'))
model.add(Dense(50, activation='relu', name='FC3'))
model.add(Dense(10, activation='relu', name='FC4'))
return model

Another well-known model is the one proposed by the team. It has an architecture similar to the NVIDIA network. I built the model and tried a few runs with different parameters. However after a few iterations in which the car had trouble completing a full lap on the track I decided to abandon it since the NVIDIA model was working OK and for the sake of delivering the project under the deadline. It may require a few iterations to adjust data and the parameters to make it work.

Image for post
Image for post network architecture

Model parameter tuning

The final model used the Adam optimizer so there was no need to tune the learning rate.

Dataset Augmentation


I scaled the pixel values of the data to be zero-centered, i.e. with a scale [-1, 1]. That is known to help neural networks to learn better and faster. As noted in the code excerpt of the NVIDIA network I chose to apply the normalization in the lambda layer of the model.

Downsampling the set

Since the car rides forward most of the time, the dataset is heavily skewed with steering angles very close to zero. For that reason I downsampled the dataset to ignore every sample where the steering angle was lower than 0.05 in either direction (left and right).

Side cameras

As suggested by the instructors I included the images provided from the left and right cameras in the training data. This was done at data load time by adding (or subtracting) an adjustment factor of 0.2 to the steering angle value capture with the centered camera.

angle_center = float(sample[3])
angle_left = angle_center + correction
angle_right = angle_center - correction

Image flip

Another strategy for augmenting the data is to add a mirrored version of a training camera image. This is useful because the training data is captured with a car moving in a counterclockwise direction, therefore most turns are left-bound. To provide the network with a more balanced number of left and right turns steering I included a flipped version of the image for every 2 out 3 samples. This also required the inversion of the steering angle.

if random() > 0.666:
img = cv2.flip(img, 1)
angle = angle * -1.0

Processing images

Different conditions of light in the simulator can affect the trained model. Sometimes a shaded area in the track shows to be tricky, and mislead the model to think the dark area is the edge of the track. In order to minimize this I augmented the dataset by randomly changing the picture brightness. This is done by converting the image to HSV color space and then applying a random multiplier to the V channel of the image.

def modify_brightness(img):
img = cv2.cvtColor(img,cv2.COLOR_BGR2HSV)
img = np.array(img, dtype = np.float64)
brightness_multiplier = 0.2
random_bright = brightness_multiplier + np.random.uniform()
img[:,:,2] = img[:,:,2] * random_bright
img[:,:,2][img[:,:,2]>255] = 255
img = np.array(img, dtype = np.uint8)
img = cv2.cvtColor(img, cv2.COLOR_HSV2BGR)
return img

I also converted the training data to the YUV color space as suggested in the original NVIDIA paper.

img = cv2.cvtColor(img,cv2.COLOR_BGR2YUV)

Another transformation in the images was to crop the top above the road where there’s mostly mountains and trees, not relevant to the training of the network. I also resized the images down to (64,64,3) which obviously speeds computation time.

Image for post
Image for post
Image for post
Image for post



My implementation uses a Python generator in order to generate data for training. This makes the program less likely to suffer memory issues during training as the data is never fully loaded at once in the computer memory.

Training and Validation data

I used a ratio of 80:20 to split the data between training and validation sets.

train_samples, validation_samples = train_test_split(samples, test_size=0.2)train_generator = generator(train_samples, batch_size)
validation_generator = generator(validation_samples, batch_size)

Driving with the model

In order to test the trained model I used the code of provided by Udacity. In order to make it work I had to modify the code to include the same image augmentation methods that were used in training (i.e. color channel conversion, cropping and resizing.)

Autonomous navigation

This is the video of a successful ride around track 1

And this is a video of the ride seen in the Udacity simulator.

Written by

Multi time founder, CTO and product executive. Founder @ridehugobus, EIR: @human_ventures & @techstars. VC: @collabfund

Get the Medium app

A button that says 'Download on the App Store', and if clicked it will lead you to the iOS App store
A button that says 'Get it on, Google Play', and if clicked it will lead you to the Google Play store