The effect of the choice of features on unsupervised clustering in audio surveillance is investigated. The importance of individual features in a larger feature set is first analyzed by examining the component loadings in principal component analysis (PCA). The individual sound events are then assigned into clusters using the self-tuning spectral clustering and the classical K-means algorithms. A weighted version of the original set is used, where the weights have been optimized by a genetic algorithm (GA) for maximally error-free clustering. The weighted feature set expectedly outperforms the original feature set and its PCA-reduced version. Insight into the importance of individual features is also gained.
Click to purchase paper as a non-member or login as an AES member. If your company or school subscribes to the E-Library then switch to the institutional version. If you are not an AES member and would like to subscribe to the E-Library then Join the AES!
This paper costs $33 for non-members and is free for AES members and E-Library subscribers.