Steering Angle Prediction in Autonomous Vehicles Using Deep Learning

The enhancement in the development of the computer vision and deep learning has led to a renaissance in the automation domain. These advancements can be applied to the work of driving a vehicle where drivers from the car can be replaced by such technology. Most important in designing the autonomous vehicle is to predict the steering angle by which the vehicle should take a curve based on curvature of the road. In this paper, the modules are split into image preprocessing and convolutional neural network. For each point along the trajectory, the steering angle at that point is recorded along with the instantaneous speed at which the car is travelling, and a set of 3 images left; centre; right captured by onboard forward-facing cameras in the vehicle and other corresponding parameters like throttle and brake. The first phase considered is image preprocessing which provides a variety of images assisting the training phase. This variation is achieved using the techniques of data augmentation which includes translation, resize, crop, flip of images. The significance of preprocessing is to eliminate any unwanted features of the image and enhance the things relevant to the model. The second phase of convolutional neural network is the deep learning phase where the preprocessed images from the previous phase is the input and steering angle is predicted as an output. For the accurate prediction of the steering angle it uses multilayer convolutional neural network and multilayer perceptron model. The method of supervised deep learning helps to validate the predicted steering angle and its accuracy. The model achieves strong performance as generic features and are even more effective when fine-tuned to large datasets. Finally, this predicted steering angle is fed to the Udacity’s simulator of driverless car to move the vehicle along the track with all its attributes.