# Real-Time Low-Cost Traffic Monitoring Based on Quantized Convolutional Neural Networks for the CNOSSOS-EU Noise Model

## TL;DR

This paper introduces a low-cost, real-time traffic monitoring system using a Raspberry Pi and a quantized CNN to accurately classify vehicles for urban noise modeling.

## Contribution

The novel contribution is a cost-effective, edge-based vehicle recognition system optimized for the CNOSSOS-EU noise model using a quantized YOLOv8 CNN.

## Key findings

- The system achieves 14 FPS inference speed and 92.2% mean Average Precision on a Raspberry Pi with Coral TPU.
- In real-world testing, it outperformed a commercial solution with a 6.6% error rate versus 59.9%.
- The system bridges the gap between manual counting accuracy and automated efficiency for traffic monitoring.

## Abstract

Accurate urban noise mapping requires granular traffic flow characterization aligned with specific acoustic models, such as CNOSSOS-EU. Existing monitoring solutions often lack the specific categorization capabilities, cost-effectiveness, or flexibility required for large-scale deployment in resource-constrained environments. To address this challenge, the present study describes the development of a real-time multi-vehicle recognition system based on low-cost edge computing hardware, specifically a Raspberry Pi 4 coupled with a Coral TPU accelerator. The proposed methodology integrates a quantized YOLOv8 convolutional neural network (CNN) with a tracking algorithm to enable real-time detection and classification of vehicles into five distinct classes, allowing for precise aggregation according to CNOSSOS-EU standards. The model was trained on a proprietary dataset of 15,000 images and subjected to 8-bit post-training quantization to optimize inference speed. Experimental results demonstrate that the system achieves an inference speed of 14 FPS and a mean Average Precision (mAP@50) of 92.2% in daytime conditions, maintaining robust performance on embedded devices. In a real-world case study, the proposed system significantly outperformed a commercial traffic monitoring solution, achieving a weighted percentage error of just 6.6% compared to the commercial system’s 59.9%, effectively bridging the gap between manual counting accuracy (1.4% error) and automated efficiency.