Self-Driving Car Simulation

Project Overview

This deep learning project implements an autonomous vehicle control model using behavioral cloning. By training on dashboard camera feeds and corresponding human steering actions recorded inside a simulator environment, the neural network learns to predict optimal steering angles dynamically to keep the vehicle centered on a winding course.

Key Features & Objectives

• Behavioral Cloning Pipeline: Maps camera image streams directly to continuous steering control parameters.
• Real-time Image Augmentation: Preprocesses image states dynamically during training to handle variations in lighting, shadows, and recovery angles.
• End-to-End Deep Learning: Bypasses manual road marker mapping, extracting steering features directly from raw image inputs.

Technical Stack & Technologies

• Language: Python
• Frameworks: TensorFlow, Keras
• Computer Vision: OpenCV (image processing)
• Simulator Platform: Udacity Self-Driving Car Simulator

Core Techniques & Deep Architectures

• NVIDIA CNN Architecture: Implemented a stacked deep convolutional neural network mirroring the NVIDIA autonomous vehicle setup (5 convolutional layers for spatial feature extractions followed by 4 fully connected dense layers).
• Camera Angle Recovery Offset: Incorporated left and right camera perspective offset calculations (adding directional bias) to train the model to recover if the vehicle drifts toward road edges.
• Data Augmentation Pipeline: Designed real-time image augmentation filters (random translations, brightness modulations, horizontal flips, and cropping) to prevent overfitting to the course layout.