AI Sound Event Detection for Road Safety in Tunnels

Use Case

The Challenge

A major operator of the national road network has launched a pilot project to enable tunnels to “listen”: to automatically recognize, based on audio from microphones, sounds that signal danger (braking, collisions, screams, car horns, sirens, tire blowouts) and alert the operations center. The main challenge was twofold: the lack of real-world recordings of tunnel accidents and a hostile acoustic environment—dominated by ventilation, traffic, and reverberation—in which critical events must be reliably distinguished with as few false alarms as possible.

The Solution

The project involved the development of a Sound Event Detection solution based on deep learning models for audio. A data preparation engine “set the scene” in the tunnel for 297 event sounds, blending them with real tunnel noise and simulating their echo: this resulted in a dataset of 2,970 samples across 8 classes (crashes, braking, horns, screams, etc.). A comparison of three pre-trained architectures (EfficientAT, AST, CLAP) allowed us to select the model with the best balance between accuracy, false alarms, and computational efficiency. The model analyzes the audio in 5-second windows and sends alerts to the operations center; a web app allows for the management of rules, thresholds, and monitoring.

The Benefits

During testing, the model achieved an accuracy of over 90%, with good performance even for the rarest events. Calibrating the thresholds cut false alarms in half compared to the alternatives evaluated. The process is repeatable and scalable: the dataset is being enriched with real-world records, and the solution is ready for field validation and extension to other tunnels.

Humanativa’s role

HUMANATIVA managed the entire lifecycle of the solution: from data preparation and signal analysis to model selection and training, and finally to the engineering of the real-time analytics service and the management web app. Thanks to the expertise of its Data Science and MLOps team, the Group delivered a documented, traceable, and reproducible solution—the first step toward an intelligent infrastructure monitoring platform.