Convolutional neural networks (CNNs) are so useful in computer vision that we are going to use one for another application, traffic sign detection for intelligent vehicles. We will also cover several important concepts of deep learning in this chapter and will get readers exposed to other popular frameworks and libraries for deep learning.

We continue our R deep learning journey with one of the core problems in self-driving cars, object recognition, and to be specific, traffic sign classification. To avoid accidents and ensure safety, robust traffic sign classification is critical to realizing driving autonomy. We will start with what self-driving cars are and what aspects deep learning is applied to. We will also discuss how deep learning stands out and becomes the state-of-the-art solution for object recognition in intelligent vehicles. With the background knowledge in mind, we'll get started with our project when we first conduct an exploratory analysis of the data and performance data preprocessing, including Y'UV color encoding conversion. We'll make our first attempt at using CNNs with MXNet and achieve more than 99% accuracy. Then we'll move on to another powerful deep learning framework called TensorFlow and an API built on top of it called Keras.

We'll also introduce the dropout technique as a regularization approach similar to Lasso (L1) / Ridge (L2) regularization. Besides overfitting, we'll demonstrate how to deal with a lack of training data, utilizing data augmentation techniques. We'll wrap up the chapter by summarizing approaches to prevent overfitting in CNN models. Once again, we will prove that deep learning models remove manual or explicit feature extraction steps that are taken in traditional approaches, and they instead automatically discover useful patterns underneath input data during the training process.

We will get into the details of the topics mentioned here:

• What self-driving cars are
• How deep learning is applied in self-driving cars
• How deep learning becomes a state-of-the-art solution for object recognition in intelligent vehicles
• Exploratory analysis on the German Traffic Sign Recognition Benchmark dataset's data
• Traffic sign recognition using convolutional neural networks with the MXNet package
• Keras and TensorFlow
• Traffic sign recognition using convolutional neural networks with the keras package
• Reducing overfitting with dropout
• Data augmentation and common augmentation methods
• Using data augmentation for a small training set
• Summarizing approaches to prevent overfitting in CNN models