Research And Implementation Of Multiple Targets Registration And Tracking System Based On Markerless Augmented Reality

In recently years,augmented reality(AR)technology has developed rapidly,its applications have gradually occupied a focus on human society.We can see AR games in daily life,such as Pokémon Go,which was very popular among teenagers.AR models can be demonstrated in the field of teaching,which give more sensory experience to students.You may also see AR conduct assemblers to install devices in the field of industrial manufacture.As technology being mature,AR applications markets get a potential prospect in the future.In early stage,marker-based augmented reality technology relied on the existence of marker template,and can't be applied into the natural feature environments,but more and more application scenes tend to adopt makerless augmented reality technology.At present,there are plenty of open software development kits provided in the market,most of them just provide rather API than the core algorithm,only ARToolKit open the source code with few research on markerless multiple targets detection and tracking.This paper takes advantage of the data structure of vocabulary tree,which is hierarchically clustering with SIFT feature point descriptors,to store the inverted files indexed by the registered targets.In detection process,we utilize the image library registered with trained vocabulary tree to retrieve multiple targets in the captured video frames,and get 3.02% improvement in retrieval performance combined with error loss value of homography transformation.In tracking process,we realize real-time multiple targets tracking by using the improved KLT target tracking algorithm,the improved KLT tracking algorithm achieves 25% improvement in precision and 37.5% improvement in success rate compared with traditional KLT algorithm.Finally,a multiple targets registration and tracking based on makerless augmented reality system was implemented.This paper comes up with makerless augmented reality approach for multiple targets registration and tracking,including online and offline stages.In offline stage,1)Extract SIFT feature point descriptors for the image sets waiting for registration,and adopt hierarchically k-means clustering to train a vocabulary tree model.2)Link an inverted file for every leaf node,and register image id whose descriptors ever occurred on the leaf node,finally,according to the length of each inverted file,a vocabulary tree with tf-idf weight is regenerated.In online stage,1)Detection thread continuously searches undetected regions for new target by using the designed multiple targets retrieval algorithm,and transfers it to a new tracking thread.2)Tracking thread makes use of improved KLT algorithm to track each target and update pose matrix,and feeds back the target status to the detect thread.3)Rendering thread update pose information with pose matrix provided by tracking thread to render 3D model.The experiment indicates that this system has a significant improvement in storage efficiency and detection speed compared with ARToolKit,and overcomes the limitation of marker-based augmented reality system,thus provides a reliable solution for markerless multiple targets augmented reality system.