🚗 Real-Time Object Detection for Autonomous Vehicles

A real-time, edge-optimized deep learning perception module for autonomous driving systems. Developed under the Digital Egypt Pioneers Initiative (DEPI) – AI & Machine Learning Track (2025).

📌 Project Overview

Autonomous vehicles require low-latency, high-accuracy environmental perception to ensure safe navigation.

This project presents a YOLOv8-based real-time object detection system optimized for edge deployment using ONNX and TensorRT, achieving:

• ⚡ 78.5 ms inference latency
• 🎯 71.1% mAP@50
• 🚘 Multi-class detection for driving-critical objects
• 🧠 Full MLOps monitoring pipeline

🧠 Architecture

🎯 Problem Statement

Autonomous systems often struggle with:

• Low-light and night driving
• Fog, rain, and weather interference
• Small object detection (traffic lights, distant pedestrians)
• Real-time latency constraints

We built a perception module that balances:

• ✅ Accuracy
• ✅ Speed
• ✅ Edge deployability
• ✅ Safety-critical robustness

🧠 Model Architecture

We selected YOLOv8 due to:

• Anchor-free detection head
• C2f backbone blocks
• Efficient gradient flow
• Native ONNX & TensorRT export
• Strong small-object detection performance

Detection Pipeline

Live Camera Feed
        ↓
Frame Preprocessing (Resize 640x640)
        ↓
YOLOv8 Inference
        ↓
NMS & Post-processing
        ↓
Structured Detection Output

📊 Dataset Strategy

📦 Datasets Used

• KITTI (Autonomous Driving Focused)
• COCO (Transfer Learning Base)
• OpenImages (Edge Cases & Diversity)

🔄 Data Engineering

• Multi-dataset fusion
• Class balancing
• YOLO format conversion
• Scene-leakage prevention
• Annotation normalization

🌧️ Custom Augmentation Pipeline

• Night simulation
• Fog & rain injection
• Motion blur
• Mosaic augmentation
• Brightness/contrast shifts

🏋️ Training Configuration

Parameter	Value
Model	YOLOv8
Image Size	512
Batch Size	16
Optimizer	AdamW
LR	0.0032
Epochs	80
Mixed Precision	Enabled

🔬 Experiments Conducted

• Baseline
• Augmentation testing
• Hyperparameter tuning (Optuna)
• Image size benchmarking
• Fine-tuning strategy

Tracked with MLflow.

📈 Performance Results

Metric	Result
Inference Latency	78.5 ms
mAP@50	71.1%
FPS (Edge Optimized)	35–45 FPS
Target Latency	< 200 ms (Exceeded by 2.5× faster)

Key Observations

• +12.7% accuracy gain after extended tuning
• Strong generalization in urban scenarios
• Remaining challenge: small traffic light detection

⚙️ Deployment & Optimization

Model Export

• PyTorch (.pt)
• ONNX (.onnx)
• TensorRT Engine (.engine)

Edge Optimization

• FP16 quantization
• Threaded inference pipeline
• Hardware-aware tuning
• Jetson validation

🔄 MLOps Integration

Built full monitoring pipeline:

• Model versioning (MLflow)
• Dataset versioning (DVC)
• Drift detection
• FPS & latency monitoring
• Automated retraining triggers
• Confidence distribution tracking

🏗 Project Structure

├── data/
├── notebooks/
├── src/
│   ├── training/
│   ├── inference/
│   ├── deployment/
│   └── mlops/
├── models/
├── experiments/
├── exports/
│   ├── onnx/
│   ├── tensorrt/
├── demo/
└── README.md

🎥 Live Demo

📺 YouTube Demo: https://youtu.be/UQaHwTuYU6g

🚀 Future Work

• Upgrade to YOLOv8-Medium
• Improve small-object detection
• INT8 Quantization
• Multi-sensor fusion (LiDAR + Camera)
• Real vehicle hardware integration

🧪 Tech Stack

• Python
• YOLOv8 (Ultralytics)
• PyTorch
• OpenCV
• ONNX
• TensorRT
• MLflow
• DVC
• Optuna

👨‍💻 Team

Developed as part of Digital Egyption Pioneers Initiative AI Cohort 2025:

• Elsayed Ali Elsayed Aboulila
• Nizar Hossam Hussein
• Abd El-Rahman Ahmed
• Ahmed Ashraf Abbas
• Mohamed Ashraf Mohamed

Supervised by: Dr. Sherif Salem

🏁 Conclusion

This project goes beyond model training - it delivers a deployable, monitored, edge-ready perception module designed for safety-critical autonomous systems.

It demonstrates full-stack ML engineering: From data engineering → model optimization → deployment → monitoring.