Date of Publication :25th June 2024

Abstract:The developed Convolutional Neural Network (CNN) for a simulation-based self-driving car project adopts an innovative end-to-end approach, mapping raw pixel data from a single front-facing camera directly to steering commands. This approach has proven highly effective in autonomously navigating through diverse road scenarios, including local roads without lane markings and highways, as well as challenging environments like parking lots and unpaved roads. The CNN autonomously learns internal representations of key processing steps, such as detecting relevant road features, using the human steering angle as the sole training signal. Unlike traditional methods that decompose the problem into explicit components, the end-to-end system optimizes all processing steps concurrently, leading to superior performance and more compact systems. The approach not only enhances overall system efficiency by self-optimizing internal components but also allows for the development of smaller networks, minimizing the number of processing steps needed for effective self-driving functionality.

Reference :