Reproduce “Audio-Based Pedestrian Detection in the Presence of Vehicular Noise” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yonghyun Kim, Chaeyeon Han, Akash Sarode, Noah Posner, Subhrajit Guhathakurta, Alexander Lerch
Researchers from Georgia Institute of Technology developed ASPED v.b, a new dataset for audio-based pedestrian detection in traffic-heavy urban soundscapes, showing that training models with vehicular noise significantly improves their ability to distinguish pedestrians from vehicles and reduces false positives. The study confirmed that models learn complex acoustic features beyond simple loudness for detection.
Key insights
- Audio-based pedestrian detection models exhibit limited domain generalization, with performance degrading when tested on acoustic environments different from their training data.
- Training models with exposure to vehicular noise is crucial for learning to differentiate pedestrian sounds from vehicle sounds, leading to a significant reduction in false positives for vehicle-related audio.
- Models rely on more complex acoustic features than simple loudness for detection, but may misclassify certain musical instruments and show reduced confidence in human sound detection in noisy environments due to masking effects.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.