Research On Human Target Detection Method Based On Hazeing Processing Algorithms

Fire accidents have been one of the problems that society attaches great importance today.Due to the difficulies in eliminating fires,it is necessary to do a good job of supervision and accurate detection of fire smoke to minimize fire losses.However,smoke images obtained in complex environments often have characteristics such as low illumination,low contrast,and complex textures,making it difficult to obtain effective information.Although many scholars have done a lot of detailed research on the smoke treatment and target detection algorithm in the early stage of fire,there are still some problems,such as interference judgment due to the unsatisfactory de-fogging effect and the difficulty of search and rescue due to the low accuracy of target detection.Therefore,it is worth studying to optimize the smoke treatment algorithm to timely understand the field situation and improve the target detection model to accurately grasp the situation of trapped personnel.In this paper,computer vision algorithms and image processing and recognition related technologies are mainly used to detect the defogging problem in complex scenes.The main research contents include:(1)Image defogging processing algorithm: When the smoke concentration is low in the early stage of fire,the smoke model constructed is similar to the degradation model in dense weather,so the improved de-fogging algorithm can be used to solve the problem of smoke treatment in the early stage of fire.In order to obtain a clearer fire scene defogging image,thereby providing better data support and support for fire scene monitoring,an improved algorithm is proposed to address the problem of inaccurate atmospheric light parameter estimation in the dark channel defogging algorithm.Firstly,an adaptive algorithm is used to initially determine environmental conditions such as light intensity,and then bright and dark channel fusion algorithm is used to evaluate the global atmospheric light optimize important parameter values of atmospheric models such as transmittance.Aiming at the problem that the atmospheric physical model established by the dark channels prior algorithm can cause image distortion in the sky and pure white areas,an end-toend deep learning method is used to optimize the transmittance.The atmospheric physical model’s color attenuation,contrast,HSV difference,bright and dark channel characteristics are used as the input for deep learning,and the maxout function is used to activate it for feature extraction.The transmittance is obtained through operations such as feature extraction,multiscale mapping,convolution and pooling.Finally,curvature filtering is performed for the halo effect to achieve the goal of smoothing image details.Based on the improved defogging algorithm,the quality of fire scene images can be effectively improved,detailed and visually clear fire scene defogging images can be obtained.(2)Target detection algorithm: Aiming at the problems of low accuracy and poor realtime performance of single-stage target detection,the target detection algorithm YOLOv5 s is improved,CBAM attention mechanism is added to the backbone network to better extract features of feature space and channel,and neck network is improved,GSConv convolutional module is used to replace the common convolutional module.At the same time,GSConv convolution module is used to design bottleneck layer.The improved algorithm not only has a good accuracy rate,but also improves the model volume and operation efficiency.Finally,it can complete the humanoid target detection in indoor smoke environment.In summary,the algorithm in this paper can achieve smoke anthropomorphic target detection in indoor fire scenes.The experimental results show that this method is feasible and provides a reliable basis for improving the quality of fire scene rescue.