Розробка моделі вимірювання відстані з використанням підходу збільшення та модифікації архітектури YOLOV3

Determining the best model for measuring object distance, the appropriate formula for that model, and modifying the YOLOv3 architecture is the focus of this research. This was done to address the problem of object distance measurement errors using a monocular camera. In this study, the researchers used a magnification approach and modified the YOLOv3 architecture, which was then named Hybrid Dist – YOLOv3. The proposed distance measurement model does not use camera height and camera shift distance variables, so it can still measure objects that are higher than the camera height and the measurement time is faster. The only variable in the measured distance formula is the change in object image height. As for modifications to the YOLOv3 architecture, there are two types of training and test data: initial measurement data and from KITTI. The training data from the initial measurements consisted of three classes, namely person, bottle, and jerrycan, with 24, 10, and 10 samples, respectively. The detection accuracy at mAP0.50 is 0.994, 1.1, with absolute measurement error values (ɛA) of –0.274, –0.153, and –0.163. For the training data from KITTI, there are three object classes, namely pedestrian, car, and truck, with 1150, 7682, and 318 samples, respectively. From the tests conducted, the ɛA values for the pedestrian, car, and truck classes show an improvement from the previous study, which were originally 1.75, 2.49, and 4.63, to 1.37, 2.25, and 3.74. The results of this research can be applied in the automotive industry to driver assistance systems (DAS), soccer robots, or similar systems that require distance measurement