Crashzam: Sound-based Car Crash Detection

Marcin Detyniecki, Ph.D., Matteo Sammarco, Marcin DetynieckiVehits 2018 · DOI

TL;DR

This paper introduces Crashzam, a novel system for detecting car accidents using only sound captured by in-car microphones. Unlike existing methods relying on accelerometers, Crashzam offers a low-cost, hardware-agnostic solution that can be implemented on smartphones, potentially improving driver safety and response times.

What Problem Does This Paper Address?

Road safety remains a critical global concern, with millions of deaths annually attributed to car accidents. While existing automatic crash notification systems often rely on expensive, dedicated hardware like accelerometers or airbag deployment sensors, a significant portion of vehicles lack these features. Furthermore, accelerometer-based systems can suffer from false positives (e.g., from road bumps) and false negatives (e.g., rear-end collisions). This paper addresses the need for a more accessible, reliable, and cost-effective car crash detection system that can be deployed on widely available devices like smartphones.

Methods

Crashzam employs a two-model approach to detect car crashes based on audio signals:

  • Model 1: Time and Frequency Analysis: This model extracts a range of established audio features, including:
  • Zero-Crossing Rate (ZCR): Measures how often the audio signal changes sign, useful for identifying percussive sounds.
  • Signal Power: The overall energy of the signal, expected to be high for crashes.
  • Entropy: Measures abrupt changes in energy, also indicative of crashes.
  • Spectral Centroid (SC) and Spectral Spread (SS): Describe the distribution of energy across frequencies. Crashes are expected to have low SC and high SS.
  • Spectral Flux (SF): Detects changes in the spectrum between consecutive frames, useful for onset detection.
  • Spectral Rolloff (SR): The frequency below which a certain percentage of spectral energy is concentrated.
  • Spectral Entropy (SE): Measures the spread of energy across frequency sub-bands.
  • Chroma Vector (CV): Groups spectral information into 12 bins representing musical pitches.
  • Mel-Frequency Cepstral Coefficients (MFCCs): A standard representation for analyzing audio, particularly speech.
  • Model 2: Spectrogram Image Analysis: This model treats the spectrogram of the audio signal as an image. It focuses on identifying patterns of "peaks" (local maxima in amplitude) within the spectrogram.
  • Peak Vertical and Horizontal Alignment: Analyzes the distribution of gaps between peaks in time and frequency to distinguish between percussive (crash) and sustained (e.g., car horn) sounds.
  • Peak Entropy: Measures the distribution of peaks across time bins.

The final detection is a weighted combination of the outputs from these two models, aiming to leverage their complementary strengths. The system is triggered by an accelerometer detecting a significant acceleration event (over 2.5g), after which it records a buffer of audio for analysis.

Datasets

The paper introduces a novel dataset for crash sound detection:

  • AXA Winterthur Crash Campaign: Controlled crash tests conducted in Switzerland, providing 46 crash sound signals recorded from inside car cabins.
  • "Car Crash Time" YouTube Channel: Audio recordings from dashcams of real-world crashes, adding realism and including background noises like rain, hail, and screams. This contributed the majority (87%) of the crash samples.
  • Negative Samples: Sounds commonly found in car cabins (radio music, talking, engine starting, car doors) and sounds that might mimic crashes (car horn, tire skidding, harsh deceleration) were sourced from FreeSound and Urban Sound Dataset.

Overall, 410 crash sounds were collected, with 100 samples set aside for testing. The dataset is balanced between "Crash" and "Other" sound categories.

Key Insights and Ideas

  • Sound as a Reliable Crash Indicator: The core idea is that car crashes produce distinct acoustic signatures that can be reliably detected, even in noisy environments. This offers an alternative to accelerometer-based systems, which can be prone to false alarms.
  • Leveraging Smartphone Hardware: The system is designed to run on smartphones, utilizing their built-in microphones and accelerometers, making it a low-cost and widely deployable solution.
  • Complementary Models for Robustness: Combining time-frequency analysis with spectrogram image analysis provides a more robust detection system, as each model excels at identifying different aspects of crash sounds.
  • Filtering Noise with Spectrograms: The spectrogram analysis, by focusing on patterns of "peaks," inherently filters out much of the background noise and sustained sounds, improving accuracy.
  • Low Power Consumption: The paper demonstrates that activating motion sensors and the microphone on a smartphone has a manageable impact on battery life, making continuous monitoring feasible.

Key Results

  • Combined Model Accuracy: The combined Crashzam system (Model 3) achieved an overall accuracy of 0.9 on the test dataset.
  • Individual Model Performance:
    • Model 1 (Time/Frequency) achieved an accuracy of 0.875.
  • Model 2 (Spectrogram) achieved an accuracy of 0.90 (as per Figure 5b, though the text states Model 1 performs better, the ROC curves suggest Model 3 is superior).
  • Misclassifications: The most common misclassifications were with slammed door sounds.
  • Battery Impact: Activating motion sensors and the microphone on a Nexus 5 smartphone reduced battery life by approximately 50 minutes compared to idle, with the microphone consuming about 3.5% of battery capacity per hour.

Limitations and Future Work

  • Slammed Door Sounds: The system still struggles to differentiate between actual crashes and loud slammed door sounds.
  • Server Dependency (Proof of Concept): The current proof-of-concept transmits data to a server for analysis. The authors suggest embedding models directly into devices for full-time, autonomous detection.
  • Deep Learning for Spectrograms: Future work could explore using Convolutional Neural Networks (CNNs) for spectrogram image processing to discover more complex patterns.
  • Explainability: For applications in domains like insurance, the ability to explain why a crash was detected is crucial, which might require further investigation beyond the current model.
  • Other Domains: The core concept of sound-based event detection could be extended to other areas like home surveillance or health monitoring.

Summary updated

BibTeX
@inproceedings{phd2018crashzam,
  author = {Marcin Detyniecki, Ph.D. and Matteo Sammarco and Marcin Detyniecki},
  title = {{Crashzam: Sound-based Car Crash Detection}},
  booktitle = {Vehits},
  year = {2018},
  doi = {10.5220/0006629200270035},
}
Figure