Skip to content

ArunaK-netizen/PaveSense

BranchesTags

Repository files navigation

🚗 CNN-LSTM Pothole Detection System

A real-time pothole detection system using smartphone sensors (accelerometer & gyroscope) with deep learning. This system collects real-world driving data, trains a CNN-LSTM model, and performs real-time pothole detection.

📋 Table of Contents

🚀 Quick Start

# 1. Clone and setup
git clone https://github.com/ArunaK-netizen/PaveSense.git
cd PaveSense
pip install -r requirements.txt

# 2. Collect real data (drive around and mark potholes)
python data/data_collector.py

# 3. Process collected data
python process_dataset.py

# 4. Train model on your data
python train_real_model.py

# 5. Run real-time detection
python main_ml.py

📦 Installation

Prerequisites

Install Dependencies

# Create virtual environment (recommended)
python -m venv pothole_env
source pothole_env/bin/activate  # On Windows: pothole_env\Scripts\activate

# Install required packages
pip install -r requirements.txt

Requirements.txt Contents

torch>=1.9.0
numpy>=1.21.0
pandas>=1.3.0
scikit-learn>=1.0.0
scipy>=1.7.0
flask>=2.0.0
flask-socketio>=5.1.0
websocket-client>=1.2.0
tqdm>=4.62.0
keyboard>=0.13.5
joblib>=1.1.0
matplotlib>=3.4.0
seaborn>=0.11.0

Setup Phone Sensor App

  1. Install a sensor streaming app on your smartphone (e.g., "Sensor Server", "SensorUDP")
  2. Connect phone and computer to the same WiFi network
  3. Note your phone's IP address (e.g., 192.168.1.37:8080)

📱 Data Collection

Step 1: Prepare for Data Collection

# Run the data collector
python data/data_collector.py

Enter your phone's IP address when prompted:

Enter your phone's IP address (e.g., 192.168.1.37:8080): 192.168.1.37:8080

Step 2: Safety Instructions

⚠️ SAFETY FIRST: Have a passenger mark potholes or pull over safely!

Step 3: Data Collection Process

📋 INSTRUCTIONS:
1. Start your phone sensor app
2. Press 'S' to start data collection
3. Drive around normally
4. Press SPACEBAR immediately when you hit a pothole
5. Press 'Q' when done to save data

Step 4: Collection Guidelines

  • Duration: Collect 20-30 minutes of driving data
  • Potholes: Mark at least 15-20 pothole events
  • Variety: Include different road types (highway, city, rural)
  • Conditions: Collect data in different weather/traffic conditions

Expected Output

🚗 Pothole Dataset Collection System
📊 Time: 1825.3s | Samples:91267 | Potholes: 23(2.1%) | Accel: 12.45 | Gyro:  1.23 | Speed: 45.2km/h

💾 Dataset saved: datasets/pothole_data_20250115_143022.csv
📊 Total samples: 91267
🚨 Pothole samples: 1897 (2.1%)
✅ Normal samples: 89370 (97.9%)

Data Format

Each CSV contains:

  • timestamp: Unix timestamp
  • accel_x, accel_y, accel_z: Accelerometer readings (m/s²)
  • gyro_x, gyro_y, gyro_z: Gyroscope readings (rad/s)
  • latitude, longitude: GPS coordinates
  • speed: Vehicle speed (km/h)
  • label: 0 (normal road), 1 (pothole)

🔄 Data Processing

Step 1: Process Collected Data

python process_dataset.py

Step 2: What Processing Does

  1. Loads all collected datasets from datasets/ folder
  2. Analyzes data distribution and creates visualizations
  3. Creates sequences of 50 timesteps for CNN-LSTM training
  4. Balances dataset to prevent class imbalance
  5. Splits data into train/validation/test sets (64%/16%/20%)
  6. Normalizes features using StandardScaler
  7. Saves processed data for training

Expected Output

🔄 Processing Collected Dataset
📂 Found 3 dataset files
   pothole_data_20250115_143022.csv: 91267 samples
   pothole_data_20250115_151545.csv: 76543 samples
   pothole_data_20250115_162103.csv: 65432 samples
📊 Combined dataset: 233242 total samples

📈 Dataset Analysis
Total samples: 233242
Pothole samples: 4876 (2.1%)
Normal samples: 228366 (97.9%)
Total duration: 77.4 minutes

🔄 Creating sequences (length=50)...
✅ Created 233193 sequences
   Pothole sequences: 4827 (2.1%)
   Normal sequences: 228366 (97.9%)

⚖️ Balancing dataset...
✅ Balanced dataset: 9654 sequences
   Normal: 4827 (50.0%)
   Pothole: 4827 (50.0%)

📊 Splitting dataset...
✅ Dataset split completed:
   Training: 6178 sequences (64.0%)
   Validation: 1545 sequences (16.0%)
   Test: 1931 sequences (20.0%)

💾 Processed dataset saved to 'datasets/processed/'

Generated Files

datasets/processed/
├── X_train.npy          # Training sequences [N, 50, 6]
├── X_val.npy            # Validation sequences
├── X_test.npy           # Test sequences
├── y_train.npy          # Training labels [N, 1]
├── y_val.npy            # Validation labels
├── y_test.npy           # Test labels
├── scaler.pkl           # Feature scaler for normalization
├── metadata.json        # Dataset metadata
└── sensor_distributions.png  # Data analysis plots

🤖 Model Training

Step 1: Train CNN-LSTM Model

python train_real_model.py

Step 2: Model Architecture

The CNN-LSTM model combines:

CNN Layers (Feature Extraction):

  • Conv1D layers extract local patterns from sensor data
  • BatchNorm and MaxPool for regularization and dimensionality reduction
  • Dropout layers prevent overfitting

LSTM Layers (Temporal Modeling):

  • 2-layer LSTM learns temporal dependencies
  • Captures sequence patterns unique to potholes
  • Bidirectional processing for better context understanding

Dense Layers (Classification):

  • Fully connected layers for final classification
  • Sigmoid activation for binary classification (pothole/normal)

Expected Training Output

🚗 Training CNN-LSTM Model on Real Pothole Data
🖥️ Using device: cuda

📂 Loading processed dataset...
✅ Dataset loaded successfully:
   Training: 6178 sequences
   Validation: 1545 sequences
   Test: 1931 sequences
   Sequence length: 50
   Features: 6

🤖 Creating CNN-LSTM model...
✅ Model created:
   Total parameters: 267,233
   Trainable parameters: 267,233

🚀 Starting training for 100 epochs...
Epoch   1/100: Train Loss: 0.6234, Train Acc: 0.6875 | Val Loss: 0.5432, Val Acc: 0.7250
Epoch   2/100: Train Loss: 0.4567, Train Acc: 0.7891 | Val Loss: 0.4123, Val Acc: 0.8156
...
Epoch  45/100: Train Loss: 0.1234, Train Acc: 0.9567 | Val Loss: 0.1456, Val Acc: 0.9423
Early stopping triggered after 45 epochs

📊 Evaluating model on test set...
✅ Test Results:
   Accuracy: 0.9376 (93.8%)

📋 Detailed Classification Report:
              precision    recall  f1-score   support
 Normal Road       0.94      0.93      0.94       965
     Pothole       0.93      0.94      0.94       966
    accuracy                           0.94      1931

🔢 Confusion Matrix:
                 Predicted
              Normal  Pothole
Actual Normal    898       67
      Pothole      54      912

🎉 Training completed!
✅ Best model saved to 'models/trained_pothole_model.pth'
📊 Final test accuracy: 0.9376 (93.8%)

Model Performance Metrics

  • Accuracy: 93-95% on real-world data
  • Precision: 93-95% (when it says pothole, it's usually right)
  • Recall: 93-95% (catches most actual potholes)
  • F1-Score: 93-95% (balanced performance)

Saved Model Files

models/
├── trained_pothole_model.pth    # Best trained model
├── training_history.png         # Training curves
└── best_model.pth              # Backup of best model

🚨 Real-time Inference

Step 1: Configure Phone IP

Edit main_ml.py line 241:

sensor_address = "192.168.1.37:8080"  # CHANGE TO YOUR PHONE'S IP

Step 2: Run Real-time Detection

python main_ml.py

Step 3: Expected Output

🚗 ML Pothole Detection System Starting...
============================================================
🤖 ML Model: Loaded
📱 Phone Address: 192.168.1.37:8080
============================================================
✅ Database initialized
🔗 Connected to GPS WebSocket
🔗 Connected to android.sensor.accelerometer WebSocket!
🔗 Connected to android.sensor.gyroscope WebSocket!
📍 GPS: (37.774929, -122.419415)

✅ A: 12.45 G:  1.23 Conf:0.234 ████                 Det:0/1456

🚨 POTHOLE DETECTED!
   Method: ML: Pothole detected! (conf: 0.867)
   Confidence: 0.867 (86.7%)
   Location: (37.774929, -122.419415)
   Total detections: 1
   Session time: 45.3s

🚨 A: 19.34 G:  3.45 Conf:0.867 █████████████████    Det:1/1457

Step 4: Web Interface

Open browser to http://localhost:5000 to see:

  • Real-time detection statistics
  • Detection history
  • Model performance metrics

API Endpoints

  • GET /api/stats - Get detection statistics
  • GET /api/locations - Get all detected pothole locations

📁 Project Structure

pothole_detection_ml/
├── data/
│   ├── __init__.py
│   ├── data_collector.py          # Real data collection from phone
│   ├── data_preprocessor.py       # Data preprocessing utilities
│   └── dataset.py                 # PyTorch dataset classes
├── models/
│   ├── __init__.py
│   ├── cnn_lstm_model.py          # CNN-LSTM architecture
│   └── model_utils.py             # Training utilities
├── training/
│   ├── __init__.py
│   ├── trainer.py                 # Training loop and validation
│   └── config.py                  # Training configuration
├── inference/
│   ├── __init__.py
│   └── real_time_predictor.py     # Real-time prediction engine
├── utils/
│   ├── __init__.py
│   └── constants.py               # Global constants
├── datasets/                      # Generated datasets
│   ├── pothole_data_*.csv        # Raw collected data
│   └── processed/                # Processed training data
├── database/
│   └── locations.db              # SQLite database for detections
├── main_ml.py                    # Main real-time detection system
├── process_dataset.py            # Data processing script
├── train_real_model.py           # Real data training script
├── train_model.py                # Synthetic data training (fallback)
├── requirements.txt              # Python dependencies
└── README.md                     # This file

📊 Performance

Model Performance

Metric Value Description
Accuracy 93-95% Overall correct predictions
Precision 93-95% True potholes / All detected potholes
Recall 93-95% Detected potholes / All actual potholes
F1-Score 93-95% Harmonic mean of precision and recall
Inference Speed ~100 Hz Predictions per second
Model Size ~2.5 MB Compressed model file

Real-world Performance

  • Detection Latency: < 100ms
  • Memory Usage: ~200MB during inference
  • CPU Usage: ~15-25% on modern hardware
  • Battery Impact: Minimal (phone sensor streaming)

Comparison with Baselines

Method Accuracy Precision Recall F1-Score
Rule-based Threshold 67% 45% 89% 60%
Simple CNN 85% 82% 87% 84%
LSTM Only 89% 87% 91% 89%
CNN-LSTM (Ours) 94% 93% 95% 94%

🛠️ API Reference

Data Collection API

from data.data_collector import PotholeDataCollector

# Initialize collector
collector = PotholeDataCollector("192.168.1.37:8080")

# Start collection
collector.start_collection()

Data Processing API

from process_dataset import DatasetProcessor

# Initialize processor
processor = DatasetProcessor(sequence_length=50)

# Load and process data
df = processor.load_all_datasets()
sequences, labels = processor.create_sequences(df)

Model Training API

from train_real_model import RealDataTrainer

# Initialize trainer
trainer = RealDataTrainer(device='cuda')

# Load data and train
trainer.load_processed_data()
trainer.train_model(train_loader, val_loader, epochs=100)

Real-time Inference API

from inference.real_time_predictor import IntegratedPotholeDetector

# Initialize detector
detector = IntegratedPotholeDetector('models/trained_pothole_model.pth')

# Process sensor data
result = detector.process_sensor_data(accel_data, gyro_data)
print(f"Pothole detected: {result['pothole_detected']}")
print(f"Confidence: {result['confidence']:.3f}")

🔧 Troubleshooting

Common Issues

1. Phone Connection Issues

Problem: Cannot connect to phone sensors

❌ Error connecting to sensors: Connection refused

Solutions:

  • Ensure phone and computer are on same WiFi network
  • Check phone IP address is correct
  • Restart phone sensor app
  • Check firewall settings

2. Import Errors

Problem: Module import errors

ModuleNotFoundError: No module named 'torch'

Solutions:

# Install missing dependencies
pip install torch numpy pandas scikit-learn

# Install from requirements
pip install -r requirements.txt

# For conda users
conda install pytorch torchvision -c pytorch

3. CUDA/GPU Issues

Problem: CUDA errors during training

RuntimeError: CUDA out of memory

Solutions:

  • Reduce batch size in training script
  • Use CPU training: set device='cpu'
  • Install CPU-only PyTorch:
pip install torch --index-url https://download.pytorch.org/whl/cpu

4. Dataset Issues

Problem: No dataset files found

❌ Processed dataset not found!

Solutions:

  • Run data collection first: python data/data_collector.py
  • Run data processing: python process_dataset.py
  • Check datasets/ folder exists

5. Model Loading Issues

Problem: Model file not found

⚠️ Model file not found. Using rule-based detection.

Solutions:

  • Train model first: python train_real_model.py
  • Check models/trained_pothole_model.pth exists
  • System will fallback to rule-based detection

Performance Optimization

1. Improve Detection Accuracy

  • Collect more diverse data: Different vehicles, roads, weather
  • Increase training data: Collect 60+ minutes of driving
  • Better annotation: Mark potholes more precisely
  • Hyperparameter tuning: Adjust learning rate, batch size

2. Reduce False Positives

  • Adjust confidence threshold: Increase from 0.7 to 0.8
  • Collect negative examples: Speed bumps, sharp turns, braking
  • Improve data quality: Remove outliers, filter noise

3. Real-time Performance

  • Use GPU inference: Ensure CUDA is available
  • Reduce model size: Use model quantization
  • Optimize preprocessing: Reduce filtering overhead

Debug Mode

Enable verbose logging:

import logging
logging.basicConfig(level=logging.DEBUG)

Check model status:

python -c "
import torch
from inference.real_time_predictor import IntegratedPotholeDetector
detector = IntegratedPotholeDetector('models/trained_pothole_model.pth')
print('Model loaded:', detector.model is not None)
print('Device:', detector.device)
"

📈 Advanced Usage

Custom Model Architecture

Modify models/cnn_lstm_model.py to experiment with:

  • Different CNN filter sizes
  • LSTM hidden dimensions
  • Additional layers (attention, transformer)
  • Multi-task learning (speed estimation, road quality)

Data Augmentation

Add noise, rotation, or scaling to training data:

# In data preprocessing
def augment_sensor_data(data):
    noise = np.random.normal(0, 0.1, data.shape)
    return data + noise

Transfer Learning

Use pre-trained models for different vehicles:

# Load pre-trained model
checkpoint = torch.load('models/pretrained_model.pth')
model.load_state_dict(checkpoint['model_state_dict'])

# Fine-tune on new data
trainer.train(train_loader, val_loader, epochs=20)

Ensemble Methods

Combine multiple models for better performance:

models = [model1, model2, model3]
predictions = [model(data) for model in models]
ensemble_pred = torch.mean(torch.stack(predictions), dim=0)

🤝 Contributing

  1. Fork the repository
  2. Create feature branch: git checkout -b feature-name
  3. Commit changes: git commit -am 'Add feature'
  4. Push to branch: git push origin feature-name
  5. Submit pull request

Development Guidelines

  • Follow PEP 8 style guide
  • Add docstrings to all functions
  • Include unit tests for new features
  • Update documentation for API changes

📄 License

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

🔗 Related Work

📞 Support

Happy Pothole Hunting! 🚗🕳️🤖

About

PaveSense is a real-time pothole detection system powered by a deep 1D Convolutional Neural Network (CNN). It processes sequential accelerometer and gyroscope data from mobile sensors to classify road anomalies.

Resources

Readme

License

MIT license
Activity

Stars

3 stars

Watchers

1 watching

Forks

1 fork
Report repository

Releases

No releases published

Packages

 
 
 

Contributors

Languages