Space Moving Object Detection And Tracking Based On Deep Learning

Moving Object detection and tracking is a popular research area in computer vision, which has numerous applications in many domains,including intelligent transportation, video surveillance, visual reality, and public safety. Consequently, it is an important research area in the computer vision field as it also promotes the development of other computer vision research areas. However, the complexity of the environment and the uncertainty of the moving object bring great challenge to the moving target detection and tracking task. The traditional detection and tracking methods based on manual feature extraction have poor generalization ability, so that they can not meet the needs of target detection and tracking in complex moving scenes. In this thesis, deep learning is used in moving object detection and tracking to extract the high-level features of the target in order to improve the generalization ability of moving object detection and tracking methods. The main work of the thesis is as follows:(1) Aiming at the disturbance of background in moving object detection task, a moving object detection method based on deep learning and local binary similarity pattern (DSGMD) is proposed. Stacked denoising auto encoder network is trained offline to extract the high-level features of the background model and the current video frame.The features are then used to distinguish the background and the foreground which has the moving object included. By this means, the noise in the background can be filtered to a certain extent. For the patch that is distinguished as foreground, each pixel is judged by the gray value and LBSP code to determine whether it belongs to moving object or not,realizing moving object detection. Experiments show that the proposed algorithm has obvious improvement compared with the contrast algorithms in the quantitative evaluation indexs. Compared to GraphCutDiff, RMoG and MST, the F-Measure of the proposed DSGMD increases 97.4%, 36.7%, 14.6%, respectively; PWC decreases by 71.5%,55.1%, 49.9%, respectively.(2) Aiming at scale variance of moving target in visual tracking task,a visual tracking algorithm based on scare invariance and deep learning(SMS-DLT) is proposed. Stacked denoising auto encoder network is used to extract the high-level abstract features. The SURF feature which is scale invariant is integrated with the high-level abstract features to enhance the tracking robustness when the size of the object to be tracked changes significantly. In the process of motion estimation, the result of particle filter is verified and corrected by mean shift to further improve the accuracy of moving target tracking. Experiements show that the proposed SMS-DLT obtains better tracking result than the compared tracking algorithms. Compared with DLT, the average center location error of the proposed SMS-DLT reduces 8.2 in pixel, and success rate increases 27.6%; Compared with So-DLT, the average center location error of the proposed SMS-DLT reduces 10.6%, and success rate increases 3.8%.(3) Aiming at the challenges such as scale variation, illumination changes and deformation of the target in the tracking tasks, a visual tracking algorithm based on mixed features (SoH-DLT) is proposed,considering both the contour features and detail features. Orientation histogram is introduced to describe the contour features of candidate samples because of it insensitivity to scale and illumination change and deformation. Deep neural network is used to extract high-level abstract features, and it is fine-tuned by SURF feature. Thus, SURF and high-level features are integrated to describe the detail of candidate samples. The experimental results show that proposed SoH-DLT has a better tracking performance than the contrast algorithms in both quantitative and qualitative evaluation. The average center location error of the proposed SoH-DLT reduces 3.6 in pixel and success rate increases 20.8% compared to DLT which does not introduce orientation histogram.(4) A moving object detection and tracking system based on deep learning is designed and implementated. The main function modules include video sequence preprocessing module, moving object detection module and moving object tracking module. The main functions of video sequence preprocessing module are gray processing, true value labeling and so on. The main functions of moving object detection module are real-time moving object detection and detection effect evaluation. The main functions of moving object tracking module are real-time tracking and tracking effect evaluation. The evaluation indexs of moving object detection mainly include recall rate, accuracy rate, error rate and so on.The evaluation indexs of visual tracking mainly include center error,coverage rate, success rate and so on. The system verifys the effectiveness of the proposed algorithm.