Research On Dynamic Object Tracking Algorithm Of Complex Scene In Coal Mine

At present,natural disasters in coal mine in our country are still serious,and coal mine safety guarantee capabilities are still relatively low.In recent years,our country has spent a lot of manpower and material resources to upgrade the coal mine production monitoring system,but manual operation has limited application to the coal mine safety production monitoring system,the efficiency is low and the error rate is high.The use of computers for intelligent tracking and monitoring of dynamic object can greatly reduce the waste of manpower,improve monitoring efficiency,and provide more reliable and concise safety early warning and system linkage than manual operation.At present,our country has ushered in a new round of industrial transformation,technological innovation is the primary productive force,and dynamic object tracking can further contribute to our country's industrial intelligence.In addition,object tracking can provide information support for behavior understanding,event detection,and object classification,and its importance is self-evident.The object tracking algorithm has achieved good results in open scenes with rich features and sufficient illumination.However,in confined spaces such as coal mine and tunnels,due to unfavorable factors such as coal dust,lack of illumination,occlusion,scale changes,and image blurring,etc.The stability and robustness of tracking are greatly affected.Aiming at the above problems,this paper improves two dynamic object tracking algorithms suitable for complex scenes in coal mine based on the generation model method and discriminant model method of visual tracking,and improves the object detection algorithm.The main tasks are as follows:(1)Multi-feature fusion dynamic object tracking algorithm for complex scenesTo solve the problem that the traditional Mean Shift tracking algorithm is easy to lose the target in scenes such as target deformation,occlusion or lighting changes in the coal mine,adopt principal curvature and color features to jointly establish the object tracking model: the second-order Gaussian filter is used to extract the texture information of the tracking object in different directions,the Hessian matrix is used to calculate the main curvature of the image surface through the texture information,and the principal curvature and RGB color information are combined to establish the object feature probabilistic histogram model.establish a Mean shift algorithm that integrates multiple features to replace the optical flow method in the TLD(Tracking-Learning-Detection)framework,reducing the computational complexity of the tracking module,and use the PN learning strategy to build a fast cascaded detector,combined with bhattacharyya coefficient to achieve accurately detect tracking failures and reinitialize the tracker to quickly corrects the tracking result.The dynamic object tracking experiment is carried out on the public datasets and coal mine monitoring video.The experimental results show that the improved algorithm can achieve long-term tracking of dynamic object in complex scenes,and has high robustness and tracking accuracy.(2)Kernelized correlation filtering tracking algorithm based on deep feature fusion for complex scenesThe KCF(Kernel Correlation Filter)tracking algorithm uses HOG features to train correlation filter,which cannot cope with the problems of object scale changes and object occlusion in underground coal mine.In this paper,the deep network model Image Net-VGG16 is used to extract the deep features of the tracking object,and the deep features and the HOG feature position filter calculation results are weighted merged to obtain the final object position,which improves the tracking stability and robustness of KCF algorithm under insufficient illumination and scarce feature points in the coal mine.In addition,a scale filter is added to enable the tracking algorithm to adapt to changes in target size.When the object is occluded,in order to prevent the filter from being contaminated,the APCE(Average Peak-to-Correlation Energy)value is introduced to detect whether it is occluded.If it is occluded,the position filter template update is stopped,and the scale filter update is stopped at the same time to avoid tracking drift and failure caused by scale changes.The dynamic object tracking experiment is carried out on the public datasets and coal mine monitoring video.The experimental results show that the improved algorithm is more robust than the comparison algorithms and can meet the needs of object tracking in complex scenes in coal mine.(3)Moving object detection and object tracking in coal mineThe object tracking algorithm runs on the public datasets and relies on the calibration of the object in the datasets.In practical applications,the tracking object cannot be calibrated in advance.Therefore,the accuracy of moving object detection is very important,and its detection effect directly affects the follow-up tracking accuracy.After analyzing the respective characteristics of the frame difference method,optical flow method,and background difference method,the background difference method is selected to detect moving object in coal mine.Aiming at the shortcomings of the background difference method,such as the error detection of shadows and light disturbances,the frame difference method is used to update the model in the background difference method to improve the robustness of the background model,and propose a method of adaptive adjustment of model learning rate.At the same time,Canny edge detection is used to further suppress the shadow of the moving object to get more accurate moving object detection results.Experiments on dynamic objects detection and tracking in coal mine show that the improved algorithm can correctly detect moving objects to lay foundation for accurate tracking.Compared with the original algorithm,the improved object tracking algorithm in this paper has higher stability and robustness,and meets the needs of object tracking in complex scenes in coal mine.Combined with the improved background difference method,it can reduce the error caused by the moving object detection,and realize the automatic detection and tracking of dynamic object.The thesis has 72 graphs,8 tables,and 88 references.