Research And Application Of Target Location And Tracking Based On Deep Learning

In recent years,algorithm of deep-learning-based object detection has achieved significant results.However,for video target tracking,the algorithm is not able to use the relationship between the timing-sequence of video,once the target detection algorithm is applied to each frame of video.For tracking tasks,this causes problems such as drift and large errors,and the target detection algorithm usually cannot meet the real-time performance.The research in this thesis begins with the algorithm requirements of the "YaMeiLe doctor-patient communication software platform".In order to solve the problem of poor texture and high similarity of the tracking target(teeth)and the problem of high real-time algorithm,this thesis proposes a deep learning-based Siamese network target tracking framework for teeth target.This thesis has carried out research work on tracking algorithms for teeth target.The main work includes:1.This thesis proposes a video key frame extraction algorithm combined with multifeature analysis.The thesis use the anonymous customer oral video collected by "YaMeiLe doctor-patient communication software platform" as the source data,and use this key frame extraction algorithm for key frame extraction and annotate data manually personally.This dataset is the first tracking and segmentation dataset of color tooth images(not CT).The extraction algorithm effectively solves the problem of removing redundant frames of video data that does not improve network performance under the condition of limited labor costs,which greatly reduces the pressure of manual labeling.2.For the problem of multiple and repeated labeling of video sequences,this thesis conduct deep study of the optical flow network and apply it to the key frame labeling data set,which effectively solved the problems of over-fitting the network due to the small data set and non-convergence of the model.3.This thesis study the feature extraction network and solve the problem that the tracking algorithm based on the Siamese framework cannot use deeper networks.Then the thesis propose an improved feature extraction network ResNet41.The FPN technology is added to the network to improve the ability to express features.And added a two-stage network to enhance the network's ability to discriminate.And increase the Mask prediction branch,further improve the tracking accuracy.Finally,the effectiveness of the above modules was verified through horizontal comparison experiments.4.This thesis use the deep learning-based Siamese network target positioning and tracking framework proposed in this thesis to design and develop “YaMeiLe doctorpatient communication software platform”.It can be applied in doctor-patient communication of stomatology and beauty related surgery in stomatology hospitals,dental clinics and other places.It can quickly collect video through the camera of mobile device,do the work of simple editing with fast background calculation to obtain the postoperative operation of dental surgery effect image and video.