Video Based Fire Detection Algorithm Research

With the development of social and economic, ecological environment is being deteriorated and the forests are shrinking, while forest fires further eroding the precious green resources. Of course, the fire hasn't only been a natural disaster, but also existed in people's daily life, and it has seriously done harm to people's lives and property. Therefore, over the years scientists have been studying fire protection and monitoring. In the early days, the fire detectors are based on the sensors. In recent years, with the computer research in the field of video processing, video-based fire detection systems have developed rapidly.Existing fire detection algorithms can be roughly classified into following categories:(1) color space based algorithms, which detect the flame area based on the flame characteristics in the color spaces, such as RGB and HIS; (2) statistical based algorithms, which detect the flame based on the probability distribution (e.g., Gaussian Probability distribution) of the flame area; (3) lifting wavelet algorithm based methods; and (4) clustering based algorithms, and so on. With respect to flame region detection, researchers mainly focus on the following regional characteristics: regional growth characteristics of early fire, the regional center features and edge features and so on. With regard to the fire verification stage, researchers mainly utilize neural networks, genetic algorithms and support vector machines. This thesis first reviews some typical fire recognition algorithms, and then proposes two novel methods for video-based fire detection:one is based on VQ (Vector Quantization), and the other is based on flame contours (edges).Traditional machine learning-based fire recognition schemes are with high complexity, hence it cannot fulfill low false alarm rate and high detection speed simultaneously and cannot assess the severity grade. In order to overcome these shortcomings, this thesis proposes a novel fire recognition algorithm based on VQ. First, we generate a fire codebook and a non-fire codebook offline by the LBG algorithm based on the training set that are selected from 10 video clips under different scenes and conditions. And then we merge the two codebooks into one codebook and sort the codewords in the ascending order of their mean values for the future Equal-average Equal-variance Equal-norm Nearest Neighbor Search (EEENNS) based fast encoding process. In the online fire detection process, the video to be detected is first segmented into successive frames, and then we perform the VQ encoding process to find fire-colored frames and record the grade of each fire-colored area. Then the moving pixel detection process is performed on each fire-colored frame to find candidate fire frames. Finally we verify whether a fire occurs or not and grade the fire by analyzing the change in the number of blocks belonging to each grade between consecutive frames. Experimental results demonstrate the effectiveness of the proposed scheme, and a 93.3% detection rate is obtained with 25 test video clips.Edge information is one of the main features of an object, and thus it has been widely used in various pattern recognition areas. Based on the phenomena that the fire edge feature changes as the fire spreads, we propose a novel contour-based early flame detection algorithm. It includes three parts, i.e., candidate fire frame selection, flame region selection, and flame contour-based fire decision. In the first step, the suspicious frames are detected and the unlikely frames are removed based on frame selection rules. The second step detects the flame pixels in the candidate fire frames by flame region selection rules. In the last step, four operations (i.e., dilation, erosion, mini region erasing and Canny edge detection) are performed on all flame regions to obtain the exact flame contours, and then fire decision rules based on three characteristics (i.e., area, perimeter and roundness of flame contours) are employed to determine whether a fire occurs in the video or not. The proposed approach is tested with several video clips in different environments, and the experimental results demonstrate its effectiveness.