🚗 Autonomous Self-Driving Car using Udacity Simulator

📌 Project Overview

This project presents the development of an autonomous driving model using deep learning techniques, specifically Convolutional Neural Networks (CNNs), trained on driving behavior data collected from the Udacity Self-Driving Car Simulator. The objective is to predict steering angles from camera images in real time, enabling the vehicle to drive autonomously within the simulated environment.

The approach emphasizes:

• End-to-end deep learning for behavioral cloning.
• Data augmentation to increase robustness.
• Efficient training pipeline to minimize overfitting.
• Real-time simulation testing using the Udacity simulator.

---

🚀 Model Training Summary

📥 1) Data Loading & Preprocessing — Robust & Reproducible

The project follows the Udacity-style dataset with driving_log.csv and three camera streams:

• Center / Left / Right camera frames are loaded and paired with steering values.
• Path Normalization: path_leaf() ensures consistent filename handling across systems.
• Steering Map: every image is mapped to its steering measurement so the model learns direct image → action mappings.

---

📊 2) Data Balancing & Distribution Control — Smart & Effective

Driving data is heavily biased toward “straight” frames. We fix that with a statistically sound approach:

• Visualize steering distribution into 25 bins.
• Cap each bin at 400 samples to avoid learning a “always-straight” policy.
• Shuffle and resample to keep training diverse.

Impact: Balanced training data improves turn detection, reduces bias, and produces safer steering behavior.

---

🎨 3) Data Augmentation Pipeline — Real-World Robustness

On-the-fly augmentation increases dataset variety without extra storage:

• Horizontal Flip — mirror images and invert steering (angle → -angle).

• Brightness Variation — HSV-based random brightness simulates day/night/shadows.

• Camera Offset Correction — left/right frames use steering offsets to teach recovery behavior.

• Random Shifts (x/y) — simulates lateral drift and camera jitter.

• Crop & Normalize — remove sky/hood and normalize pixels to [-1, 1] for stable training.

  

Why this is strong: augmentation simulates realistic edge cases (glare, shadow, slight off-center) so the DNN generalizes to new roads.

---

🧠 4) Model Architecture — NVIDIA-Style CNN (DNN)

A compact, production-minded Deep Neural Network tuned for steering regression:

• Preprocessing: Cropping + normalization layers inside the model for portability.
• Convolutional Backbone: 5 Conv layers with ReLU to capture low→high level visual features (lanes, curbs, shadows).
• Fully Connected Head: Dense layers with dropout for robust decision-making.
• Output: Single linear neuron predicting the continuous steering angle.

Design goal: real-time-capable, small enough for deployment, powerful enough for complex road geometry.

---

⚙️ 5) Training Configuration — Stable & Reproducible

Training follows best-practice defaults and production controls:

Component	Setting
Optimizer	Adam
Loss	Mean Squared Error (MSE)
Batch size	64
Epochs	5–10 (adjust per validation curve)
Callbacks	EarlyStopping, ModelCheckpoint

Generator-based training: Augmentation is applied inside a generator for memory-efficiency and infinite variety.

---

📈 6) Training & Evaluation — Track, Visualize, Validate

• Loss vs Epochs charts to monitor convergence and detect overfitting.
• Qualitative tests: Overlay predicted steering on sample frames for visual sanity-checks.
• Quantitative checks: MSE on validation set and sample-driven error analysis.

---

🏁 7) Final Output — Production-Oriented Results

The trained model:

• Predicts steering angle from a single camera frame.
• Handles lighting variation, lane curvature, and small lateral drifts.
• Produces smooth, stable steering suitable for simulator deployment.

---

📁 Project Structure

├── Self_Driving_Car.ipynb       # Jupyter Notebook for training
├── model.h5                     # Trained Keras model
├── driving_log.csv              # Driving behavior log
├── IMG/                         # Folder with camera images
├── utils/                       # (Optional) For helper functions
└── README.md                    # Project documentation

---

🚀 Getting Started

Prerequisites

• Python 3.7+
• TensorFlow / Keras
• OpenCV
• NumPy, Matplotlib, Pandas
• Udacity Self-Driving Car Simulator

pip install -r requirements.txt

Running the Model

1. Launch the Udacity simulator.
2. Use the trained model (model.h5) with your inference script.
3. Drive autonomously using predicted steering angles.

---

📷 Images & Demo

---

---

👨‍💻 About Me

Hi, I’m Rohith Boppana — a passionate and driven final-year B.Tech student in Computer Science and Engineering with a specialization in Artificial Intelligence & Machine Learning.

I'm deeply interested in building real-world tech solutions that combine data, intelligence, and intuitive design. My academic journey and hands-on projects reflect a strong foundation in both theory and practical application.

👇 My Core Interests

• 🤖 Artificial Intelligence & Machine Learning
• 🔍 Data Science & Analytics
• 📊 BI Dashboards & Predictive Modeling
• 💡 Problem-Solving with Scalable Technologies

I enjoy translating business needs and data insights into impactful software solutions that solve real problems and enhance user experiences.

---

🔗 Let’s Connect

📫 LinkedIn
Let’s connect and grow professionally:
linkedin.com/in/rohith-boppana-39ab57279

🌐 Portfolio
Explore my latest work, skills, and projects here:
rohith-boppana.vercel.app

---

💡 “Final-year student, forever learner — building the future, one project at a time.”

Feel free to explore my repositories and reach out for collaborations, internships, or to discuss innovative ideas!