Applying Techniques in Supervised Deep Learning to Steering Angle Prediction in Autonomous Vehicles

The development of effective autonomous vehicles is, in today’s day and age, an undoubtedly popular application of research in machine learning and control. One of the most fundamental tasks in the design of algorithms to control autonomous vehicles is steering angle prediction based on realtime driving data. In this paper, we discuss the application of techniques in supervised learning to predict the vector of steering angles required to navigate a car through some specified trajectory. The attributes upon which we construct features and train supervised learning algorithms are provided in a dataset recently released by Udacity [1]. Each file in the dataset describes a trajectory (roughly a 15 to 45 minute drive), over the duration of which attributes are collected. For each point along the trajectory, the steering angle at that point is recorded along with [1] the GPS coordinates of the car, [2] the instantaneous speed at which the car is traveling, and [3] a set of 3 images {left,center,right} captured by onboard forward facing cameras in the vehicle. Our analysis of the problem can be split broadly into three phases, with primary emphasis on the last phase. The first phase concerns the prediction of steering angle from GPS coordinates and speed alone. The second phase of the analysis concerns the evaluation of conventional techniques in supervised learning (such as support vector regression) to predict steering angle from (1) pre-processed dataset images and (2) extracted features of these images. The third and final phase of the analysis concerns the evaluation of deep learning schemes to predict the steering angles (from both raw images and pre-processed ones). For this last phase in particular, we consider both a multi-layer convolutional neural network and a multilayer perceptron model.