A thesis project evaluating the feasibility of using Convolutional Neural Networks (CNNs) to classify and count vehicles based on audio input, all running on cost-efficient microcontrollers.
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.
| Test Type | Accuracy | F1-Score | Counting Accuracy |
|---|---|---|---|
| Software-based | 84.8% | 87.0% | - |
| Lab-based MCU | 88.4% | 90.4% | 99.8% |
- Name: IDMT-traffic
- Classes:
car,motorcycle,commercial vehicle(bus/truck),background_noise - Source: Jakob Abeßer, Saichand Gourishetti, András Kátai, Tobias Clauß, Prachi Sharma, Judith Liebetrau IDMT-Traffic: An Open Benchmark Dataset for Acoustic Traffic Monitoring Research, EUSIPCO, 2021.
- Device: Raspberry Pi Pico 2W
- Constraints: Memory and compute optimized
- Implementation: Audio framing (0.25s), soft plurality voting for temporal accuracy
Bachelor's Programme in Information and Communication Technology
School of Electrical Engineering and Computer Science, KTH Royal Institute of Technology