🏷️ BS Thesis: Emergency Vehicle Detection with Deep Learning 🚨

This repository contains the BS Thesis focused on automatic detection of emergency vehicles from audio signals using deep neural networks in Python.

📋 Link to paper:

---

📂 Project Structure

├─── data # Datos con los que entrenar los modelos
│    ├─── csv_files  # Etiquetas en CSV
|    |     ├─── balanced_train_segments.csv
|    |     ├─── class_labels_indices.csv
|    |     ├─── eval_segments.csv
|    |     └─── unbalanced_train_segments.csv
│    ├─── mfcc.pkl
│    ├─── lfcc.pkl
│    └─── chroma.pkl
├─── models # Modelos obtenidos tras entrenar (MODELO_FEATURE_LOSS.keras)
│    ├─── cnn_chroma_0.1107.keras
│    ├─── cnn_lfcc_0.5374.keras
│    ├─── cnn_mfcc_0.2168.keras
│    ├─── ff_chroma_0.3828.keras
│    ├─── ff_mfcc_0.5162.keras
│    ├─── lstm_chroma_0.2633.keras
│    ├─── lstm_lfcc_0.6681.keras
│    ├─── lstm_mfcc_0.3320.keras
│    └─── encoders # Encoders obtenidos para reducción de la dimensionalidad
│         ├─── encoder_chroma.keras
│         ├─── encoder_lfcc.keras
│         └─── encoder_mfcc.keras
└─── src # Archivos para entrenar
     └─── eda.ipynb
          ├─── evaluation.ipynb
          ├─── download_data  # Archivos para la descarga de los audios desde YouTube
          │    ├─── dna.ipynb
          │    └─── dpa.ipynb
          ├─── feature_extraction # Archivos para obtener las características de los audios
          │    ├─── chroma_extraction.ipynb
          │    ├─── lfcc_extraction.ipynb
          │    └─── mfcc_extraction.ipynb
          └─── training # Archivos para entrenar los autoencoders y los modelos de clasificación
               ├─── autoencoders.ipynb
               └─── models.ipynb

---

🛠️ Requirements

• Python ≥ 3.11
• Install dependencies:

  pip install -r requirements.txt

---

🦜Demostration usage

1. Run backend

   cd backend
   python api.py

2. Run frontend

   cd frontend
   netlify dev

🚀 Technical details

1. Data Download & Preprocessing: Audio files have already been downloaded and preprocessed. Relevant code can be found in:

   • src/download_data/dpa.ipynb
   • src/download_data/dna.ipynb

   For positive and negative audio files respectively.

   You should work with features extracted from: src/feature_extraction. They are stored in:

   • data/mfcc.pkl
   • data/lfcc.pkl
   • data/chroma.pkl

   Format for all three cases:

   feature = {
      'positive': [],
      'pnames': [],
      'negative': [],
      'nnames': []
   }

   In positive, values for positive data is stored, while in pnames their .wav names are stored (same index). Same thing for negative and nnames.

   A detailed exploratory data analysis is available in: src/eda.ipynb.

2. Model training:

   Autoencoders can be trained in src/model/training/autoencoders.
   You should select the desired technique at the top of the file.

   Classification models can be trained in src/model/training/models.
   You should choose the architecture and technique you wish to use at the top of the file.

3. Evaluation & Metrics:

   python src/evaluation.ipynb

   You should choose the architecture and technique you wish to use at the top of the file.

---

Ángel Morales – Computer Engineering student - University of Cadiz