Research On Detection And Recognition Of Video Moving Target In Underground Coal Mine

The environment of coal mine underground production,is complex and harsh,disaster accidents always threaten the safety of national assets and personnel.The study of mine safety production has become an important topic.In this paper,the detection and recognition of moving targets in underground video surveillance are studied,and the detection and recognition methods of moving targets with better performance and effect are proposed.(1)The quality of video images in underground coal mine monitoring is not high,and the limitations of light,noise,shadow and other conditions make the single algorithm not good for image detection.The paper proposes an improved metjhod to solve the problem of poor detection effect of downhole video images:The background model is established by the mixed Gauss method,foreground image is detected by the background difference method,and fusion of the foreground image detected by the inter-frame difference method and processed Canny edge detection operator are used to detect the filled moving object mass.Then,using different logic operat:ion to achieve the target detection.The detection result obtained by the fusion is optimized by removing the shadow and connected domain identifier.Incorporating the edge detection operator,which is insen,sitive to illumination and noise,so it can eliminate the multi-light and noise in the image.Although the hybrid gaussian background modeling has good robustness,the background update speed may not keep up with the changing speed of the moving object,the background update of the three-frame difference method is intended to complement this defect.The experimental results show that the method effectively improves the detection effect of video images in underground coal mine.(2)Combining with the traditional target detection results and deep learning algorithm,and mine video target recognition method based on RFCN(Region-based Fully Convolutional Networks,RFCN)algorithm,improves the recognition accuracy of traditional convolutional neural networks and the phenomenon of over-fitting.Through the analysis and research of the network model in deep learning and the deep learning algorithm of RFCN,the Google ResNet101 network structure is adopted to achieve the network level of 101 layers to ensure the target recognition accuracy.Moreover,the residual mechanism is adopted in this scheme,without increasing the running time.Firstly,the network model is trained experimentally to achieve satisfactory results.Then,the causes of missed detection and false detection in the identification are analyzed.The secondary training is performed on the difficult samples,and the parameters of the network model are fine-tuned to obtain the optimal detection and identification model.The experiment was based on the Windows platform,the Python development language was used for programming,and the tensorflow open source framework was used for simulation training.Firstly,the experiment proves that the detection effect of the fusion classic target detection algorithm in the coal mine underground video image is improved;Then,the recognition and positioning effect of moving objects contained in the underground mine monitoring video is confirmed,and other video images in underground mine are further used for testing.It has higher recognition accuracy and good robustness,fast running time and good real-time performance,which solves the problem that it is difficult to balance the accuracy and real-time performance of the deep network structure.