Trafikljudsövervakning med neurala nätverk på mikrokontroller : En studie om klassificering och räkning av fordon

This study evaluates the feasibility of classifying and counting bypassing vehicles based on audio using Convolutional Neural Networks (CNNs) on a cost-efficient Microcontroller Units (MCUs). The classification task involved distinguishing four classes: car, motorcycle, commercial vehicle, and background noise. A lightweight CNN was trained on extracted Mel-frequency cepstrum coefficients (MFCCs) (audio features) from a pre-recorded dataset. Software-based tests were conducted to see if the model could perform well without the MCU, while lab-based tests were done after the model was deployed on the MCU. The software-based test used TensorFlow to evaluate the classification rate. In contrast, the lab-based test used a program built to simulate an audio stream to the MCU alongside a classification voting process to compensate for short audio frames. The results showed a classification accuracy of 84.8% and F1-score of 87.0% during the software-based test. While the lab-based test showed a higher classification accuracy of 88.4% and F1-score of 90.4%. The lab-based test also showed a vehicle counting accuracy of 99.8%. The study’s goal was to investigate the feasibility of using CNNs on-board MCUs for classifying and counting passing vehicles. By following the guidelines in the report, the results confirm that this is indeed a feasible alternative to other vehicle counting alternatives. It also serves as a foundation for using classification in real time.