Sound Detection and Localization in Windy Conditions for Intelligent Outdoor Security Cameras

Abstract

Sound event detection and localization (SDL) is helpful for extracting information about the position of sound sources in real time using a microphone array. This paper develops an SDL system for intelligent outdoor security cameras, so that it can listen and react to the surrounding acoustic events. In outdoor environments, this task is challenging due to high-energy and non-stationary noises such as wind noise. This paper proposes new methods for improving both detection and localization, based on a new feature, namely cross-channel power difference (XPD). The XPD is estimated from the difference of short-term power between microphones that are sensitive to wind noise. In the detection step, a time frame with high XPD is regarded as wind noise, and periods of wind, which cause false alarms, are removed from the localization step. Furthermore, the XPD is used to create a binary mask for separating the wind noise and other sound sources, thus preventing the wind noise from degrading the localization of target sounds. The proposed system is evaluated using a hardware prototype that consists of four microphones attached to the housing of a pan-tilt-zoom camera. Through real environmental experiments, we indicate that the proposed methods outperform other state-of-the-art SDL methods in windy conditions.