Smoke Recognition And Location Algorithm Based On Multi-feature Fusion

Fire early warning can reduce the losses caused by disasters.In most cases,smoke rather than open fire appears and spreads first in fire.Therefore,smoke detection and automatic and accurate identification can realize fire early warning in time.The traditional monitoring methods may be constrained by the size of the region and affected by the surrounding environment,or if the smoke concentration is low,the range is small,and it is not easy to detect the smoke,the accuracy and robustness of the detection algorithm are difficult to meet the expectations.These can overcome the shortcomings of previous detection methods by improving the smoke detection algorithm.How to distinguish smoke and things similar to its characteristics and how to extract smoke characteristics quickly and accurately have become the breakthrough point of smoke detection.The work and innovations of this thesis are as follows:Firstly,aiming at the shortcomings that the traditional fusion algorithm of local binary pattern(LBP)and Gabor features is not suitable for smoke detection,a smoke detection algorithm of TDFF is proposed to make up for the above defects.Firstly,T-MFLBP model is established,which can capture clearer texture features according to all the gray changes between pixels and can be combined with the relationship characteristics between pixels in special positions in non-uniform mode;At the same time,the first derivative of Gaussian function is obtained and used as a new filter to extract wavelet features,so as to improve the performance of capturing edge information;Finally,the fusion feature matrix is sent to support vector machine for training,and the training model can be used for smoke recognition.Experimental results show that TDFF algorithm has high detection rate and robustness for smoke detection,and has certain advantages.Secondly,aiming at the shortcomings of low detection rate and low detection rate caused by the redundancy of feature number extracted by traditional Haar method and poor robustness after fusion with LBP,a smoke detection algorithm of WHFF is proposed to make up for the above defects.For WHFF fusion algorithm,firstly WRLBP model is proposed,which extracts smoke features by optimizing the relationship between central pixels and neighborhood pixels to improve robustness;The Haar-like feature template is further improved and the gray level is compressed,so that the Haar template removes redundant features and reduces the computational complexity;Finally,the fused feature matrix is sent to the classifier for classification.Experimental results show that WHFF algorithm greatly improves the real-time performance and detection rate.Thirdly,in view of the shortcomings of traditional smoke detection technology,such as real-time,poor reliability,unable to distinguish between background and smoke,resulting in low recognition rate,G-CGBlob video smoke detection and location algorithm is proposed.This algorithm improves the gray level co-occurrence matrix,obtains the training model combined with support vector machine,and then uses the CGBlob proposed in this thesis to convert the RGB image of video into frame into HSV color space,and delimits the threshold for three color parameters respectively;Then the RGB image is converted into gray image at the same time,and the threshold value is set for the gray value and the adjacent gray change range.Using the above methods,the exact smoke position in the image can be detected,and finally the detected area can be located by binocular camera.Experimental results show that the algorithm can accurately recognize the smoke image and locate it in time.Finally,this thesis summarizes the advantages and shortcomings of the improved algorithm,and looks forward to the future development prospect.