Safe Drive is a next-generation machine learning system that transforms the driving experience. Our project addresses the growing need for enhanced safety, luxury, and automation in modern vehicles, leveraging intelligent systems to provide a safer, more connected driving experience.
This vision aligns with Qualcomm’s leadership in automotive AI through their Snapdragon Ride™ and Snapdragon Elite™ platforms, paving the way for smarter, connected cars. Inspired by this blog, Safe Drive integrates cutting-edge AI models to enhance the overall driving experience.
Driving isn't just about reaching your destination; it's about how safely and comfortably you get there. With Safe Drive, we’ve integrated advanced AI models to tackle challenges like drowsiness detection, low-light driving, and voice-based automation, creating a system that truly cares for the driver.
💡 Built for Real Needs: Modular design for enhanced flexibility and future updates.
💡 AI at Its Core: Real-time detection, prediction, and decision-making powered by cutting-edge models.
SafeDriveAI25.mp4
Download the video from the drive: https://drive.google.com/file/d/16CmlSPQNxpy2xsP9rfr9S4vJ3fG4EAB5/view?usp=sharing
Safe Drive integrates multiple AI models, many of which rely on object detection to ensure safety and performance. Below is an overview of its features and how object detection is utilized:
- 🔒 Verifies the driver using voice biometrics.
- Object Detection: Uses feature detection to identify unique voice patterns.
- Feature Extraction: MFCC for unique audio characteristics.
- Model: Siamese Network with a ResNet-18 backbone and attention mechanism for one-shot learning.
- 📁 Folder:
Voice_Authentication/
- 👤 Recognizes the driver and opens a personalized profile built with Django and frontend languages.
- Object Detection: Detects the driver’s face using a KNN-based recognition system.
- 📁 Folder:
Facial_Recognition/
- 🗨 Allows hands-free operation through voice commands.
- Feature Extraction: Embeddings generated via Word2Vec.
- 📁 Folder:
Speech_Recognition/
- 🛡 Prevents fatigue-related accidents using live video analysis.
- Object Detection: Detects eyes and facial features with YOLOv5 for real-time fatigue assessment.
- 📁 Folder:
Drowsiness_Detection/
- 🚧 Identifies obstacles and assists in route navigation.
- Model: Built with YOLOv5 for robust and accurate object detection.
- 📁 Folder:
Object_Detection_And_Navigation/
- 🌌 Enhances visibility during nighttime or low-light conditions.
- Model: Uses the MIRNet model from Hugging Face for low-light image enhancement.
- 📁 Folder:
Low_Light_Vision/
- ⚠ Recognizes and interprets road signs in real time.
- Object Detection: Utilizes YOLOv5 for high-speed detection of traffic signs.
- 📁 Folder:
Traffic_Sign_Detection/
- 🚗 Enhances road safety by dynamically controlling headlight beams to prevent glare.
- Object Detection: Detects vehicles using a trained TensorFlow model.
- Hardware Integration: Arduino-controlled LED matrix for adaptive lighting.
- 📁 Folder:
Advanced_Headlight_System/
Each model has been rigorously tested to ensure reliability in real-world scenarios. Below are the models with their respective architectures and accuracies:
| Feature | Model/Architecture | Accuracy |
|---|---|---|
| Voice Authentication | Siamese Network (ResNet-18 Backbone + Attention) | 85% |
| Facial Recognition | KNN-based System | 96% |
| Speech Recognition | Word2Vec Embeddings | 92.5% |
| Drowsiness Detection | YOLOv5 | 90% |
| Object Detection & Navigation | YOLOv5 | 89.5% |
| Low Light Vision | MIRNet from Hugging Face | 75% clearence |
| Traffic Sign Detection | YOLOv5 | 96% |
git clone https://github.com/prabhuanantht/SafeDriveAI.git
cd SafeDriveAIFurther steps are explained inside the folders of each model.
- Resource Usage:
- Memory: ~3-4GB RAM
- Disk Space: ~24GB for training
- Training Time:
- On Google Colab TPU: ~20-30 minutes for 10 epochs
- Inference Speed:
- ~5-10ms per file for most modules
- Maximum latency: ~10 seconds during complex inference
- Performance measured using:
- Precision: Ensures correct predictions for minimal errors.
- Recall: Effectively captures true positives in predictions.
- SSIM (Structural Similarity Index): Evaluates the quality of image enhancements in models like Low Light Vision.
- PSNR (Peak Signal-to-Noise Ratio): Measures the image enhancement performance to ensure high-quality outputs.
- Designed for real-time processing:
- Object detection and navigation systems achieve latency under 10ms per frame.
- Driver authentication systems respond in less than 1 second.
- Trained on a diverse, large-scale dataset, enabling:
- High adaptability to real-world scenarios.
- Effective handling of various environmental conditions, such as low light, cluttered backgrounds, and diverse driver profiles.
- Combines multiple cutting-edge models to enhance:
- Driver safety: Prevent accidents with proactive alerts.
- Automation: Hands-free operation with advanced speech and object recognition.
- Luxury: Personalized profiles, voice authentication, and intelligent vision systems.
- Scalable design to support:
- Larger datasets with increased computational resources.
- Integration with future AI-powered features for next-gen automotive tech.