Abstract:
This thesis aimed at employing IoT and TinyML techniques to develop a system that accurately detects vehicle crash accidents using acoustic data and triggers an emergency alert in real time to the nearest police unit. The methodology involved collecting secondary data on vehicle crashes, training a machine learning model, and deploying it on the Arduino Nano 33 BLE microcontroller, which supports TinyML. This technology enables low-power, on-device machine learning for embedded systems, facilitating real-time data processing and decision-making while extending battery life without requiring cloud connectivity. To detect a vehicle crash, the model uses acoustic data to identify an accident and transmits the GPS location via a LoRaWAN communication model. The system successfully alerts and provides the exact location of the accident by showing it on google map. Results demonstrate high accuracy, with 99.3% for training and 98.4% for testing, indicating effective differentiation between accident and non-accident scenarios. When an accident is detected, the GPS location is sent to the Arduino Cloud, plotted on a map, and an alert sound is produced. Further improvements can be made by employing energy harvesting techniques, developing a dedicated mapping system, and creating a model capable of classifying and detecting various forms of accidents beyond vehicle crashes.
