Multiclass Object Classification in Autonomous Driving

Table of Contents

1. Project Overview
2. Team Members
3. Project Description
4. Dataset Links
5. Technical Details
6. Installation & Setup
7. Usage Guide
8. Results & Performance
9. Future Work
10. Contributing
11. License & Contact

Project Overview

Quick Start

git clone https://github.com/bhushanasati25/Multiclass-Object-Classification-in-Autonomous-Driving.git
cd Multiclass-Object-Classification-in-Autonomous-Driving
pip install -r requirements.txt

Project Goals

• Develop a robust object classification system for autonomous driving
• Implement and compare 8 different classification models
• Create a deployable solution with real-time inference capabilities
• Provide comprehensive model comparison and analysis

Classification Categories

1. Multi-class Classification (DenseNet-121 Specialized)
   • Human (Pedestrian, Cyclist)
   • Vehicle (Car, Truck, Van, Tram)

Team Members

Project Team

1. Bhushan Asati

   • Role: Data Scientist
   • Models: DenseNet121, MobileNetV2
   • Contributions: Data preprocessing, Feature Engineering, Model training, Model Optimiztion, Deployment

2. Rujuta Dabke

   • Role: Data Scientist
   • Models: EfficientNet, Faster R-CNN
   • Contributions: Feature engineering, Visualization

3. Suyash Madhavi

   • Role: Data Scientist
   • Models: Inception-v3, ResNet50
   • Contributions: Model optimization, Performance analysis

4. Anirudh Sharma

   • Role: Data Scientist
   • Models: XGBoost, Random Forest
   • Contributions: Traditional ML implementation, Evaluation metrics

Project Description

Project Scope

• Implementation of 8 different classification models
• Real-time inference capabilities
• Web interface for demonstration
• Comprehensive performance analysis
• Model comparison framework

Expected Outputs

1. Technical Deliverables

   • Trained model weights
   • API for inference
   • Web interface
   • Performance metrics

2. Documentation

   • Technical documentation
   • API documentation
   • User guides
   • Performance reports

Innovation Points

• Multi-model comparison framework
• Specialized DenseNet-121 implementation
• Hybrid classification approach
• Comprehensive evaluation metrics

Dataset Links

• Kaggle dataset: https://www.cvlibs.net/datasets/kitti/

• KITTI dataset: https://www.kaggle.com/datasets/garymk/kitti-3d-object-detection-dataset

Technical Details

Project Structure

Multiclass-Object-Classification-KITTI/
├── data/
│   ├── raw/               
│   ├── processed/        
│   └── samples/           
├── notebooks/
│   ├── 1_data_preprocessing.ipynb
│   ├── 2_model_training.ipynb
│   └── 3_model_evaluation.ipynb
├── models/
│   ├── resnet50_model.h5
│   ├── efficientnetb0_model.h5
│   └── ...
├── scripts/
│   ├── preprocess.py
│   ├── train.py
│   └── evaluate.py
├── streamlit_app/
│   └── app.py
└── docker/
    └── Dockerfile

Models Implemented

1. ResNet50

   • Architecture: Deep residual network
   • Features: Skip connections, Batch normalization
   • Performance: 98% accuracy

2. EfficientNetB0

   • Architecture: Compound scaling
   • Features: Balanced scaling, Optimized architecture
   • Performance: 16% accuracy

3. MobileNetV2

   • Architecture: Lightweight CNN
   • Features: Inverted residuals, Linear bottlenecks
   • Performance: 99% accuracy

4. Faster R-CNN

   • Architecture: Region Proposal Network (RPN)
   • Features: Region Proposal Network
   • Performance: 89% accuracy

5. Inception-v3

   • Architecture: Multi-scale processing
   • Features: Factorized convolutions, Auxiliary classifiers
   • Performance: 98% accuracy

6. DenseNet121

   • Architecture: Dense connectivity
   • Features: Fine-tuned for detailed classification
   • Performance: 96% accuracy

7. XGBoost

   • Type: Gradient boosting
   • Features: Feature importance, Handles imbalanced data
   • Performance: 98% accuracy

8. Random Forest

   • Type: Ensemble learning
   • Features: Feature selection, Parallel processing
   • Performance: 98% accuracy

Technical Stack

• Python 3.8+
• PyTorch
• TensorFlow
• scikit-learn
• XGBoost
• OpenCV
• Streamlit

Installation & Setup

Prerequisites

• Python 3.8+
• CUDA-capable GPU
• Git
• Docker (optional)

Environment Setup

# Create virtual environment
python -m venv venv
source venv/bin/activate  # Linux/Mac
# or
.\venv\Scripts\activate  # Windows

# Install requirements
pip install -r requirements.txt

# Download dataset
python scripts/preprocess.py --download

Dataset Setup

# Process dataset
python scripts/preprocess.py --process

Usage Guide

Training Models

# Train specific model
python scripts/train.py --model resnet50 --epochs 100

# Train all models
python scripts/train.py --all

Evaluation

# Evaluate model
python scripts/evaluate.py --model resnet50

Web Interface

cd streamlit_app
streamlit run app.py

Results & Performance

Model Comparison

Model	Accuracy	Human Precision	Human Recall	Human F1	Vehicle Precision	Vehicle Recall	Vehicle F1
ResNet50	0.98	0.88	0.99	0.93	1.00	0.98	0.99
EfficientNetB0	0.16	0.15	0.99	0.26	0.95	0.02	0.04
MobileNetV2	0.99	0.95	0.98	0.96	1.00	0.99	0.99
DenseNet121	0.96	0.77	0.99	0.87	1.00	0.95	0.97
InceptionV3	0.98	0.92	0.93	0.92	0.99	0.99	0.99
Random Forest	0.98	0.96	0.88	0.92	0.98	0.99	0.99
Faster R-CNN	0.89	0.64	0.60	0.62	0.93	0.94	0.94
XGBoost	0.98	0.96	0.94	0.95	0.99	0.99	0.99

Key Findings

1. DenseNet-121 showed best performance for detailed classification
2. Deep learning models consistently outperformed traditional ML approaches
3. MobileNetV2 provided best speed-accuracy trade-off
4. Vision Transformer showed promising results but required more training data

Future Work

1. Implement ensemble methods
2. Add real-time video processing
3. Optimize for edge deployment
4. Expand dataset with synthetic data
5. Implement cross-validation

Contributing

1. Fork the repository
2. Create feature branch (git checkout -b feature/AmazingFeature)
3. Commit changes (git commit -m 'Add AmazingFeature')
4. Push to branch (git push origin feature/AmazingFeature)
5. Open Pull Request

License & Contact

License

This project is licensed under the MIT License - see the LICENSE file for details.

Contact Information

• Bhushan Asati : [basati@stevens.edu]
• Rujuta Dabke : [rdabke@stevens.edu]
• Suyash Madhavi: [smadhavi1@stevens.edu]
• Anirudh Sharma : [asharma16@stevens.edu]

Repository

Project Link: https://github.com/bhushanasati25/Multiclass-Object-Classification-in-Autonomous-Driving.git