A Novel Vision-Based Classification System for Explosion Phenomena

Abstract

The need for a proper design and implementation of adequate surveillance system for detecting and categorizing explosion phenomena is nowadays rising as a part of the development planning for risk reduction processes including mitigation and preparedness. In this context, we introduce state-of-the-art explosions classification using pattern recognition techniques. Consequently, we define seven patterns for some of explosion and non-explosion phenomena including: pyroclastic density currents, lava fountains, lava and tephra fallout, nuclear explosions, wildfires, fireworks, and sky clouds. Towards the classification goal, we collected a new dataset of 5327 2D RGB images that are used for training the classifier. Furthermore, in order to achieve high reliability in the proposed explosion classification system and to provide multiple analysis for the monitored phenomena, we propose employing multiple approaches for feature extraction on images including texture features, features in the spatial domain, and features in the transform domain. Texture features are measured on intensity levels using the Principal Component Analysis (PCA) algorithm to obtain the highest 100 eigenvectors and eigenvalues. Moreover, features in the spatial domain are calculated using amplitude features such as the YCbCr color model; then, PCA is used to reduce vectors’ dimensionality to 100 features. Lastly, features in the transform domain are calculated using Radix-2 Fast Fourier Transform (Radix-2 FFT), and PCA is then employed to extract the highest 100 eigenvectors. In addition, these textures, amplitude and frequency features are combined in an input vector of length 300 which provides a valuable insight into the images under consideration. Accordingly, these features are fed into a combiner to map the input frames to the desired outputs and divide the space into regions or categories. Thus, we propose to employ one-against-one multi-class degree-3 polynomial kernel Support Vector Machine (SVM). The efficiency of the proposed research methodology was evaluated on a totality of 980 frames that were retrieved from multiple YouTube videos. These videos were taken in real outdoor environments for the seven scenarios of the respective defined classes. As a result, we obtained an accuracy of 94.08%, and the total time for categorizing one frame was approximately 0.12 s.