Dynamic texture based smoke detection using Surfacelet transform and HMT model

• A new dynamic texture descriptor is proposed for vision based smoke detection.
• Spatial and temporal information of image sequence are considered simultaneously.
• Coefficients of Surfacelet are modeled with 3D HMT and its trained parameters.
• Joint probability density from coefficients and parameters is feature of smoke video.

To detect smoke regions from video clips, a novel dynamic texture descriptor is proposed with Surfacelet transform and hidden Markov tree (HTM) model. The image sequence is multi-scale decomposed by a pyramid model, and the signals are decomposed to different directions using 3D directional filter banks. Then a 3D HMT model is built for obtained coefficients from Surfacelet transform with both Gaussian mixture model and scale continuity model. Parameters of the HMT model are estimated through expectation maximization algorithm, and the joint probability density is determined as the dynamic texture feature value. Support vector machine (SVM) classifier is trained with samples including smoke and non-smoke videos. For input image sequence, the joint probability density of each divided unit 3D block is taken as the input of SVM to decide whether there is smoke. The new dynamic texture descriptor takes image sequence as a multidimensional volumetric data, i.e., considering both spatial and temporal information of coefficients into one model. In experiments, existing texture descriptors of gray level co-occurrence matrix (GLCM), local binary pattern (LBP) and Wavelet are implemented and used for comparison. Results from many real smoke videos have proved that the new dynamic texture descriptor can obtain higher detection accuracy.