Deep Learning for Steering Angle Prediction in Autonomous Vehicles

Abstract

The fields of computer vision, image processing, and deep learning have seen exponential growth and advancement in recent years thanks to advances in technology. This renaissance has recently established itself as the dominant force in the field of automation. The ideas of image processing are easily applicable to the driving of automobiles, where the increasing sophistication of technology may one day make human drivers obsolete. However, the steering angle of a vehicle is the single most crucial component that comes into play when it comes to automation because this is what determines how the vehicle will take the curve. Automotive vehicles (AVs) have begun to take the steering angle prediction into consideration, and a large number of automotive businesses, like Tesla and Udacity, have also invested in the technology. Despite this, a significant number of researchers and insurance firms have shown interest in investing in this field. It has been determined that deep learning architectures are the most appropriate fundamentals that can be utilized in a situation like this one. As a result, this thesis suggests employing DL in order to anticipate the steering angle of autonomous vehicles. Image processing and CNN are the two components that make up the implementation, which is carried out in two separate modules. A collection of images are taken by the cameras that are mounted, and a steering angle is computed for each point along the path that the vehicle travels, taking into account the speed at which the vehicle is traveling as well as the amount of pressure that is being applied to the brake pedal. Image processing, which makes use of photos for the purpose of training, and data augmentation, which resizes the images, are both included in the first step of the execution process. The processed image that was obtained from the first phase is given as the input to the subsequent phase, and a predicted steering angle is generated as the output of this phase. The subsequent phase involves the use of CNN ideas.