An improved yolov4-tiny-lite object detection model for mobile augmented reality

Abstract

Augmented Reality (AR) is a technique that overlays digital information on objects or places in the real world to enhance the user experience. In particular, the AR technique is promising to provide unprecedented immersive experiences in the fields of entertainment, marketing, education, industry, fashion, and healthcare. Most existing AR techniques are able to recognize the 3D objects of the surroundings. However, they still suffer from the ability to detect complex objects in the real world. Moreover, object detection is now an important field of computer vision and AR. The goal of object detection process is to select and classify objects in many specialized fields and applications such as face detection and recognition. Several models such as YOLO, YOLO-LITE, and YOLOv4-tiny are used to detect the objects from the natural features. However, robust detection of objects from the natural features in AR development is still a complex problem and usually degraded the accuracy and requires high computational time and computation cost. To overcome these problems, this research is conducted to propose a new model based on hybridizing the You-Only-Look-Once- LITE (YOLO-LITE) with You-Only-Look- Once Version 4 tiny (YOLOv4-tiny). The proposed model is called You-Only-Look-Once Version 4-tiny-lite (YOLOv4-tinylite). Based on YOLO-LITE as the backbone network YOLOv4-tiny-lite used feature pyramid network to extract feature maps of various sizes and also use a "shallow and narrow" convolution layer to improve a detector. Thereby obtaining the ideal balance between detection speed and precision when used with portable devices and PCs without GPUs. YOLOv4-tiny-lite achieved a mAP of 52.6% for PASCAL VOC DATASET and 33.3% for the COCO dataset. YOLOv4-tiny-lite achieved about 20 FPS on a non-GPU device. The YOLOv4-tiny-lite has provided higher accuracy and less computational time compared to state of art non-GPU models. This proposed model promises to improve object detection for CPU-based devices, subsequently enhancing AR systems for mobile devices and other devices without GPUs. For future work the suggested model accuracy has to be improved in the future in order to be more accurate and useful in actual applications like autonomous driving.