An Improved YOLOv3 Object Detection Network for Mobile Augmented Reality

With the spread of mobile devices such as mobile phones, MAR(Mobile augmented reality), which is a technology that realizes augmented reality on mobile devices, is becoming one of the most popular directions in augmented reality research. In MAR, the capturing and positioning of target objects, that is, tracking and registration technology is a crucial problem. In mobile devices, tracking registration technologies that use cam-eras as tracking sensors are divided into hardware sensor-based and computer vision-based tracking registration technologies. Compared with the former, the latter has the characteristics of low hardware equipment requirements and high accuracy. However, traditional computer vision-based tracking registration technologies are susceptible to factors such as background environment, distance, and angle. To overcome the weakness, our research combines the development of deep learning in the field of object detection and lightens YOLOV3 network, which includes simplifying the network structure, improving multi-scale feature fusion detection, optimizing the dimensions of candidate frames through clustering, and optimizing the loss function, so that the object detection network can be used on mobile devices with guaranteed accuracy, and reduces the influence of background environment and other factors on the visual tracking registration technology. Our research realizes a mobile augmented reality system based on the IOS system, which achieves state-of-the-art performance.