Comparative Study of CNN Architectures for Real-Time Audio-Based Car Accident Detection on Edge Devices

Traffic accidents often result in fatalities for both drivers and bystanders. Traditionally, accident information relies heavily on community reports, which can delay the provision of victim assistance. To address this issue, a system capable of detecting accidents responsively in various weather conditions and traffic densities is necessary. One approach involved using audio analysis techniques to evaluate collision sounds. Thus, this study proposed an audio classification system for detecting car accidents using Convolutional Neural Networks (CNNs). The system’s performance was evaluated on personal computers and edge devices, such as the Raspberry Pi 4 and NVIDIA Jetson Nano, to compare inference times and power consumption. To enhance the dataset, segmentation and augmentation techniques were applied before converting the audio data into a 2D Mel-spectrogram. The dataset was then trained and assessed with four CNN architectures: custom sequential, custom with shared input layer, transfer learning EfficientNetB0, and transfer learning MobileNetV2. Both original and Lite models were deployed on experimental devices. Results showed that the custom CNN model had faster inference times across devices in both original and lite forms, though it had a 4% increase in the false positive rate. The Lite MobileNetV2 model recorded the fastest inference time on edge devices at 86 ms. Jetson Nano exhibited faster inference times compared to Raspberry Pi 4. However, Raspberry Pi 4 showed a minor increase in power consumption of 0.6 watts during inference. In future work, this system can be tested in real-time environments using embedded systems to evaluate its robustness against noise and varying environmental conditions.