Research On Human Detection In Underground Mine Videos

The production process in coal mine is complex and the environment is harsh,which seriously affects the sustainability of production and personal safety.How to strengthen production safety,improve the efficiency of early warning and rescue work are very important tasks.At present,most coal mining enterprises have deployed video surveillance systems.By detecting miners and their actions in video,the early warning and linkage control of abnormal conditions can be realized,which is of great significance for the safety production of coal mine.The mine environment has its own special features: uneven illumination,too similar gray level of miner and background in the mine video images.It is unsatisfactory to directly use traditional object detection methods to detect miners.Aiming at the characteristics of mine video,this thesis studies human detection methods,laying the foundation for subsequent analysis of abnormal behavior.The main researches are as follows:(1)Aimed at the low illumination,uneven illumination,too similar gray level of miner and background in the mine video images,this thesis puts forward a method for human detection called TD-HF.The algorithm integrates the time difference method and the human detection algorithm based on Haar features.The time difference method is not easily affected by ambient light while it can reduce the influence of low illumination and uneven illumination.On the one hand,the AdaBoost algorithm based on the Haar feature can train the classification suitable for human detection in mines.On the other hand,it can guarantee that the average time of detection can meet realtime requirements when using classifiers for detection.The experimental results show that compared with the traditional method of AdaBoost,the detection rate is increased by 8.80%,the mistake rate is decreased by 69.23% and the average detection time is decreased by 61.16%.(2)The thesis proposes a method of human detection in the mine video based on their helmet's features since the TD-HF method has defects of background block and figure mistake detection.By extracting the features of helmet's contour and color to accurately locate the human position.The experimental results show that the detection rate based on the features of the helmet is 92.72%,the mistake rate is 0.85% and the average detection time is decreased by 74.68% compared with the TD-HF method.(3)On the basis of human detection,further research on action recognition is necessary.In the thesis,the CNN-DT method is presented to recognize actions in mine video.The method introduces dense trajectory algorithm to track human detected previously and densely sample a set of points,utilizing the deep architecture to learn the multi-scale convolution feature maps of mine video at the same time.Combined with the feature maps and the dense trajectories,the deep convolutional descriptor of dense trajectory is obtained.The descriptor can represent the corresponding video.Finally,Fisher Vector is chosen to encode the deep convolutional descriptor of dense trajectory over the entire video into a global vector,the linear SVM as the classifier to perform action recognition.Experiments were conducted on five types of human actions in mine video,the experimental results show that the average recognition rate of all samples is 90.75%.Here is the basic research on action recognition in mines,which will pave the way for follow-up research on recognition of abnormal behavior.