Real-time DeepLabv3+for pedestrian segmentation

摘要

In this paper, we propose a real-time pedestrian segmentation method that is built on the structure of the semantic segmentation method DeepLabv3+. We design a shallow network as the backbone of DeepLabv3+, and also a new convolution block is proposed to fuse multilevel and multitype features. We first train our DeepLabv3+ on the Cityscapes dataset to segment objects into 19 classes, and then we fine tune it with persons and riders in Cityscapes and COCO as the foreground and the other classes as the background to get our pedestrian segmentation model. The experimental results show that our DeepLabv3+ can achieve an 89.0 mean intersection-over-union pedestrian segmentation accuracy on the Cityscapes validation set. Our method also reaches a speed of 33 frames per second on images with a resolution of 720 x 1280 with a GTX 1080Ti graphics processing unit. Experimental results prove that our method can be applied to various scenes with fast speed. (c) 2019 Optical Society of America