| Summary: | Sound event detection (SED) assists in the detainment of intruders. In recent decades, several SED methods such as support vector machine (SVM), K-Means clustering, principal component analysis, and convolution neural network (CNN) on urban sound have been developed. Advanced work on SED in a rare sound event is challenging because it has limited exploration, especially for surveillance in a forest environment. This research provides an alternative method that uses informative features of sound event data from a natural forest environment and evaluates the CNN capabilities of the detection performances. A hybrid CNN and random forest (RF) are proposed to utilize a distinctive sound pattern. The feature extraction involves mel log energies. The detection processes include refinement parameters and post-processing threshold determination to reduce false alarms rate. The proposed CNN-RF and custom CNN-RF models have been validated with three types of sound events. The results of the suggested approach have been compared with well-regarded sound event algorithms. The experiment results demonstrate that the CNN-RF assesses the superiority with remarkable improvement in performance, up to a 0.82 F1 score with a minimum false alarms rate at 10%. The performance shows a functional advantage over previous methods. © 2022, Institute of Advanced Engineering and Science. All rights reserved.
|
|---|