Research On Video Activity Analysis Methods In Underground Coal Mines

In the process of coal mine production,due to the harsh underground environment and the immaturity of related technology,processes such as support after cutting of the roadheader are basically completed by manual operation.The operation sequence and the degree of standard are directly related to the subsequent mining operations and the safety of miners.The detection and identification of personnel behavior in the video,the determination of the relationship between personnel and equipment,and the accurate prediction of coal miners’ actions,and the real-time monitoring of coal mine safety,is an academic topic worthy of study and of practical significance.This thesis conducts research on low illumination action recognition of personnel in underground coal mines,and the main research contents are as follows:(1)Low-light image enhancement in underground coal mines based on attention mechanismTo address the problems of poor image quality in coal mines due to insufficient illumination,etc.,and the difficulty of existing image enhancement techniques to reduce noise while enhancing images,This thesis proposes a low-illumination image enhancement technique in coal mines based on a hybrid attention mechanism.First,the method uses the local feature extraction branch to input the low-illumination RGB image into a local feature extraction module consisting of depth convolution and normalization layers,etc.,to achieve local feature extraction;second,it uses the global feature extraction branch to input the low-illumination RGB image into a highdimensional mapping module consisting of two convolution layers for preliminary convolution processing to obtain a high-dimensional,low-resolution RGB image.Realize image pre-processing before inputting the hybrid attention module;then pass the high-dimensional and low-resolution RGB images through feature attention mapping and pixel-level attention mapping,respectively,to extract global features;subsequently,introduce the learnable color matrix to perform matrix multiplication operations on the obtained local features and adjust the white balance parameters and CCM parameters;finally,introduce the learnable gamma parameters to extract global features after adjusting the Finally,the gamma parameters are introduced,and the gamma correction is performed on the image after adjusting the white balance and CCM parameters to obtain the final enhanced image.(2)Self-supervised learning based video action recognition frameworkFor video understanding tasks,data quality and data volume are very important.However,it is very difficult to annotate video data in the actual environment of underground coal mines,so most of the video data are unlabeled.Also,due to the very large amount of video data,acquiring high quality labeled data is time and cost consuming.To solve this problem,this thesis pre-train the model with self-supervised learning,and then fine-tune the pre-trained model on small-scale labeled video data to make the model better adapted to the task of action recognition.Technically,this approach forces the model to learn motion-related contextual information from the lowframe-rate input by predicting that subtle motion captured at high frame rates will be captured.Spatial relationships are modeled using all frames of the low-resolution view to predict the high-resolution view frames.(3)Design and application of prototype systemA prototype system for low-illumination action recognition of personnel in underground coal mines is developed based on the above theoretical foundation of the argument.Firstly,the system components are introduced.Secondly,the functions and versions of the software used in the system development and the configuration of the hardware are introduced.Finally,the operation procedures of the low-illumination enhancement function and the action recognition function are introduced.The thesis has 20 figures,4 tables and 89 references.