Acoustic Traffic Monitoring with MEMS Microphones, Concatenated log-Bark Spectrograms and CNN
TL;DR
The conclusion was that the MEMS-microphones could be used to develop edge-AI type sensors for motion’s direction determination and detection of passing vehicle in background noise using the concatenated log-Bark spectrograms converted to grayscale images as input to a CNN-classifier.
Abstract
The paper presents our results in algorithm-development and solutions for two ATM problems using MEMS microphone-signals. The aim was that the algorithms have to be realised in the sensor itself, thus we followed the edge-AI system-concept. The IDMT-Traffic dataset was used in algorithm development. The paper proposes the concatenated log-Bark spectrogram as feature extraction, followed by the custom CNN classifier. In the motion’s direction determination (from left/from right/no vehicle) the overall accuracy was 97.70±0.64%. The detection of passing vehicle in background noise was considered as a two-class classification problem (vehicle/no vehicle). The accuracy, F1-score and Matthew’s Correlation Coefficient (MCC) performance parameters achieved were 99.75%, 0.9975, 0.9949, respectively. The conclusion, based on experimental evidence, was that the MEMS-microphones could be used to develop edge-AI type sensors for motion’s direction determination and detection of passing vehicle in background noise using the concatenated log-Bark spectrograms converted to grayscale images as input to a CNN-classifier.