Road object recognition method based on improved YOLOv3

• DOI: 10.25236/ajcis.2022.050501
• Publication Date: 2022-09-14T09:22:27Z
• Authors: Yichi Zhang; Boyu Hu; Xinyi Yuan; Yupeng Li
• Keywords: Pyramid (geometry); Feature (linguistics); Realization (probability); Identification

Based on the emergence and development of autonomous driving technology, identification obstacles road is a very important challenging task. And there are many difficulties in realization this task, for example, types targets, scale span large. In view these problems, experiment proposes three improvement directions YOLOv3 algorithm to perform task target prediction: one improve up-sampling multiple use more shallow spatial information accuracy small detection. The second change way feature fusion pyramid. Thirdly, convergence direction model changed by clustering learning. Experiments BDD100K data set show that yolov3_10cls_tiny proposed paper has best detection performance better average than YOLOv3.