Modified YOLOv3-Tiny Using Dilated Convolution for Driver Distraction Detection
KyungHi Chang, Chia-Yu Wang, Chun-Han Shen · 2020
In this paper, the proposed YOLOv3-tiny-dilated network is presented. The well-known YOLO series neural network is famous for its rapidity, and the lightweight version is more suitable for mobile devices or embedded systems. However, the accuracy of one-stage network is slightly lower. Therefore, scale variation problem is investigated in this paper and we replaced the original Feature Pyramid Network (FPN) of the original YOLOv3-tiny network by three dilated convolution branches of different dilation rates to increase the accuracy of the network. As a result, the mean average precision (mAP) was improved by about 2% in the COCO dataset test. Besides, the proposed network was used to identify driving distraction behavior and obtain good accuracy of 95.47%.