Acoustic event detection for drone search and rescue system based on bi-directional long and short-term memory beamforming method to remove rotor noise

Abstract

In outdoor search and rescue operations, drone carrying microphone array to collect acoustic signals has gradually been applied to search and rescue missions for people in distress. However, since the distance between the array and the drone is much smaller than the distance between the array to the trapped person, the drone rotor noise contaminates the collected acoustic signal seriously, which often makes the search and rescue system ineffective. It is necessary to study a method that can eliminate the influence of drone rotor noise on the detection results. To address this problem, an improved Minimum Variance Distortionless Response (MVDR) algorithm based on bidirectional long short-term memory network (BLSTM-MVDR) is proposed in this paper. This method includes the bidirectional long short-term memory network to estimate the time-frequency mask of drone rotor noise, which is used as the noise covariance matrix in the MVDR algorithm, to obtain the weight vector of the algorithm for eliminating the drone rotor noise, which removes the drone rotor noise component and improves the ability of the existing algorithms to resist the drone rotor noise. Finally, a Deep Learning-based sound event detection classifier is constructed by combining convolutional neural network to realize accurate and effective search and rescue of trapped people. To verify the accuracy and effectiveness of the proposed method, a complete drone search and rescue system is constructed using the server side of the hardware device and the client side of the software platform, and the effectiveness of the system in detecting the human voice event of trapped individuals is evaluated according to different signal-to-noise ratios, sound source directions and sound source distances. The results of this paper show that the effect using the method proposed in this paper is satisfactory, with the detection accuracy improved by 43.75 % and the effective search and rescue range doubled compared with the existing methods. This method improves the accuracy of outdoor search and rescue missions in a strong drone rotor noise interference environment, and promotes the development of drone technology applications.