A novel single shot-multibox detector based on multiple Gaussian mixture model for urban fire smoke detection

Abstract

Under complex scenes, the traditional smoke detection methods cannot satisfy the real-time and accuracy requirements. Therefore, this paper proposes a novel single shot-multibox detector based on a multiple Gaussian mixture model for urban fire smoke detection. Multiple Gaussian models are used to represent the features of each pixel in the moving object image. The Gaussian mixture model is updated based on the principle that each pixel in the image is regarded as a background point if it matches the Gaussian mixture model. Otherwise, if it matches the Gaussian mixture model, it is regarded as the foreground point. By updating the foreground model and calculating the short-term stability index, the detection effect of moving objects is improved. By determining the relationship between Gaussian distribution and pixel, a new parameter is set to construct the background model to eliminate the influence caused by illumination mutation. Aiming at the problems of smoke detection efficiency and network over-fitting, we present an InceptionV3- feature fusion single shot-multibox detector. The new neural network is trained and tested by smoke positive and negative sample images. At the same time, Multibox Loss function is replaced by the Focal Loss function, which reduces the detector misdetection caused by the imbalance of positive and negative samples. Experimental results show that the proposed method is feasible and effective. The average accuracy of smoke detection is 97.5%, and the average response time of the smoke alarm is 4.57s, which can meet the requirements of real-time smoke detection in complex scenes.