Detection Model of Occluded Object Based on YOLO Using Hard-Example Mining and Augmentation Policy Optimization

• Published in: Applied sciences Vol. 11; no. 15; p. 7093
• Main Authors: Ryu, Seong-Eun, Chung, Kyung-Yong
• Format: Journal Article
• Language: English
• Published: Basel MDPI AG 01-08-2021
• Subjects: Accuracy; Algorithms; Augmentation; Autonomous vehicles; data augmentation; data mining; Deep learning; Driving ability; model optimization; Neural networks; Occlusion; occlusion detection; Optimization; Roads & highways; self-driving; Sensors; Traffic; Transfer learning
• ISSN: 2076-3417; 2076-3417
• DOI: 10.3390/app11157093

A study on object detection utilizing deep learning is in continuous progress to promptly and accurately determine the surrounding situation in the driving environment. Existing studies have tried to improve object detection performance considering occlusion through various processes. However, recent studies use R-CNN-based deep learning to provide high accuracy at slow speeds, so there are limitations to real-time. In addition, since such previous studies never took into consideration the data imbalance problem of the objects of interest in the model training process, it is necessary to make additional improvements. Accordingly, we proposed a detection model of occluded object based on YOLO using hard-example mining and augmentation policy optimization. The proposed procedures were as follows: diverse augmentation policies were applied to the base model in sequence and the optimized policy suitable for training data were strategically selected through the gradient-based performance improvement rate. Then, in the model learning process, the occluded objects and the objects likely to induce a false-positive detection were extracted, and fine-tuning using transfer learning was conducted. As a result of the performance evaluation, the model proposed in this study showed an mAP@0.5 value of 90.49% and an F1-score value of 90%. It showed that this model detected occluded objects more stably and significantly enhanced the self-driving object detection accuracy compared with existing model.