Real-Time Smoke Detection in Video Based on Two-Step Selection of Regions of Interest and Directional Movement Analysis

An algorithm for early detection of light smoke in image sequences has been developed, which enables processing of high-resolution video in real time. To do this, preliminary areas of interest that may contain smoke are identified based on motion detection, its spatiotemporal analysis, and color segmentation in HSV space. Then, the calculation of the parameters of high-frequency components is applied for these areas using a two-dimensional wavelet transform and contrast on video frames relative to the background model. This approach makes it possible to identify regions where smoke obscures background elements. The result of this stage is refined areas of interest. The final step is to estimate the direction of movement in the identified areas using the optical flow method, taking into account the analysis of changes in movement vectors over time. The algorithm is implemented using the OpenCV computer vision library in C++. Software has been developed for automated marking of video data and calculation of detection metrics. The results of experimental studies on the assessment of the accuracy and speed of the algorithm are demonstrated, and a comparison of its effectiveness with existing ones is presented.