A statistical fire judgment system for the time series of smoke density using gradient histogram

We propose a statistical fire judgment system by a new feature extraction for the time series of smoke density in an early stage. We focus on the gradient of the time series, and we quantize the time series into a gradient histogram of 32 dimensional vector. We use the gradient histogram as a feature vector, and we compared the fire judgment rates with five criteria (Euclidean distance, simple similarity, multiple similarity, Mahalanobis' distance, and modified Bayes discriminant function) as a statistical discriminant function. We evaluated experimentally the efficiency of proposed methods for actual 233 unknown time series, where the average of the correct judged rate is 60% using the conventional rule. The correct judged rates are 74% for Euclidean distance, 75% for simple similarity, 83% for multiple similarity, 83% for Mahalanobis' distance, and 85% for a modified Bayes discriminant function. By the comparison of the previous feature, it appears that the proposed feature reduced the false alarm rate from 19% to 14%.