Real-Time Horn Detection and Noise Regulation System for Silence Zones

INTRODUCTION
In many Indian cities, noise pollution caused by continuous vehicle honking has reached alarming levels—especially in silence zones like hospitals, schools, and residential areas. Traditional traffic systems are not capable of detecting or controlling unnecessary honking in real time.
Our project introduces a real-time, AI-powered horn detection system that can automatically detect, classify, and analyze honking behavior. The aim is to support noise regulation, raise public awareness, and help reduce sound pollution in urban environments.

OBJECTIVES
To develop a real-time system capable of detecting and classifying vehicle horns using Digital Signal Processing (DSP) and Convolutional Neural Networks (CNNs) for accurate audio classification.
To deploy the system on low-cost embedded hardware (ESP32) for real-world use.To categorize honking into normal, aggressive, and emergency types.
To promote quiet zones and aid enforcement of noise control regulations.

LITERATURE SURVEY
Several researchers have explored environmental sound classification and noise monitoring using advanced deep learning and signal processing techniques:
Jacome et al. (2023) used Gammatone auditory filters with Temporal Convolutional Networks (TCN) for long-term sound capture, achieving high accuracy in environmental sound localization.
Ullo et al. (2020) proposed a hybrid computerized method combining DSP and deep learning for environmental sound classification.
Shin et al. (2023) developed SELD U-Net, which jointly optimizes sound event localization and detection with noise reduction capabilities.
Toffa and Mignotte (2020) utilized Local Binary Patterns with audio feature collaboration for environmental sound classification.
Tran and Tsai (2020) demonstrated CNN-based acoustic emergency vehicle detection, highlighting CNNs' efficiency for sound-based recognition.
These studies show the evolution of audio recognition systems—from traditional DSP models to deep learning approaches. However, most existing solutions rely on cloud computation or high-end hardware. Our system stands out by performing real-time classification on an embedded microcontroller, making it cost-effective and suitable for field deployment.

METHODOLOGY
Audio Capture – The system uses an INMP441 microphone to capture environmental sounds in real time.
Signal Processing – The audio signals are processed using Short-Time Fourier Transform (STFT) to generate spectrograms, representing sound intensity across time and frequency.
Feature Extraction – The spectrograms are normalized and prepared for classification.
CNN Classification – A Convolutional Neural Network trained on a labeled horn dataset classifies each input as normal honk, aggressive honk, or siren.
Edge Deployment – The optimized and quantized model (using TensorFlow Lite) runs directly on ESP32, eliminating cloud dependency.
Output & Logging – The system stores classification results locally for further analysis or enforcement.
RESULTS AND INTERPRETATION
The CNN achieved an accuracy of over 95% on the test dataset.
Successfully distinguished between normal horns, aggressive horns, and emergency sirens despite background noise.
The embedded deployment maintained low latency (<500 ms) and efficient power usage.
The model’s lightweight design allows deployment on low-cost devices while maintaining strong
classification performance.

MERITS
Low-cost and energy-efficient solution for smart cities.
Offline operation—no need for cloud or internet connectivity.
Pivacy-preserving, as audio data never leaves the device.
Real-time performance ensures immediate detection and action.
Scalable and modular for multi-zone deployment.

DEMERITS
Accuracy can drop in extremely noisy or overlapping sound conditions.
Limited processing power on ESP32 restricts very large model use.
Requires retraining for new horn patterns or environments.
Microphone placement is crucial for accurate sound capture.

FUTURE WORK
Dataset Expansion: Collect larger and more diverse audio samples to improve model robustness.
Noise Localization: Use microphone arrays for direction-based sound detection.
Hybrid Model Integration: Combine CNN and Transformer models for better accuracy.
Edge Optimization: Implement adaptive DSP filters to handle varying noise levels.
Smart City Integration: Connect multiple devices through a network for centralized monitoring and real-time reporting.
Mobile App Interface: Develop an app for authorities to visualize honking intensity and noise patterns in different areas.

CONCLUSION
The proposed Real-Time Horn Detection and Noise Regulation System offers a promising approach to manage noise pollution in silence zones. By combining DSP and CNN on an embedded platform like ESP32, the system enables real-time, offline, and low-cost noise detection. With continued optimization and field deployment, this project can contribute significantly to creating quieter and healthier urban environments.