Vehicle-related accidents are one of the largest sources of construction worker injury. To help reduce the number of accidents involving construction workers, we introduce CSAFE, a low power wearable platform that uses multi-channel audio to detect, localize, and provide alerts about oncoming vehicles to improve construction worker safety. One of the biggest challenges in realizing this audio-based system is that construction sites are filled with hammering, power tools, and other noisy work sounds that can easily overpower the sound of an approaching vehicle. To address this challenge, we develop a novel source filtering architecture, capable of running on embedded and mobile platforms, that adaptively filters out loud construction sounds from the environment and greatly improves the detection and localization of vehicles.
Adaptive Construction Noise Filtering Architecture
To account for loud construction tools that will be present in a construction site we propose an energy-efficient sound filtering architecture that contains content-based separation and spatial separation in a feedback configuration, which utilizes known or learned models of sounds to iteratively remove noise and boost target sounds. We introduce and integrate a novel content-based separation algorithm called Probabilistic Template Matching, that is capable of running in mobile devices and leverages statistical “templates” of noises to filter out construction sounds. Our novel filtering architecture differs from existing works in sound source separation in that we intelligently leverage both single-channel source separation and multi-channel source separation in a feedback architecture to more robustly remove overpowering construction noise over time.
We incorporate our novel adaptive filtering pipeline into an embedded wearable platform consisting of an array of microphones mounted on a construction worker helmet. We leverage the embedded front-end hardware, developed in the PAWS project, to extract frequency domain features from each microphone and transmit them to the smartphone. The mobile device performs construction noise filtering using our novel adaptive filtering pipeline and runs machine learning algorithms on the filtered signals to perform vehicle detection and localization.
We evaluate CSAFE at a real construction site, with a wide variety of construction work noises, and show that CSAFE greatly improves vehicle detection and localization over existing works.
2021 | ACM/IEEE IPSN ’21
CSafe: An Intelligent Audio Wearable Platform for Improving Construction Worker Safety in Urban Environments
Stephen Xia, Jingping Nie, Xiaofan Jiang
2020 | ACM AIChallengeIoT ’20
PAMS: Improving Privacy in Audio-Based Mobile Systems
Stephen Xia and Xiaofan Jiang