An object tracking method to track maintenance objects by object detection based on deep learning in induced maintenance environment

Induced maintenance means that the surrounding scene and the maintenance object state change are perceived through natural interaction means, and the digital maintenance instruction information is superimposed on the physical object and the process in real time so that the person can obtain the operation guidance. However, in the augmented reality environment, there are few researches on application object detection technology in the field of induced maintenance. Therefore, this paper proposes a process for object detection and localization of induced maintenance based on the yolov3(you only look once)-slim model. For the problem that the deep convolutional neural network is deployed on the mobile side and the cost is too high, a pruning method for the batch normalization layer is adopted(Zoph et al. 2019). After the object is recognized, it can only obtain the two-dimensional coordinate information in the image. The coordinates obtained in the two-dimensional pixel coordinate system are then converted into three-dimensional coordinates in the camera coordinate system through the method of solving the PNP(Perspective N Points) problem(Moreno-Noguer, Lepetit, and Fua 2007). The experimental results show that the recognition rate of the repair parts based on the deep neural network can reach 90% of the mean average precision. The average recognition speed is less than 300 millisecond per frame, meeting the accuracy and interactivity requirements of the AR-induced maintenance system.