Multi-task Semantic Segmentation for Apolloscape based on CGAN

• Published in: Journal of physics. Conference series Vol. 1550; no. 3; pp. 32071 - 32075
• Main Authors: Lin, Yuankai, Yu, Xinjia, Cheng, Tao
• Format: Journal Article
• Language: English
• Published: Bristol IOP Publishing 01.05.2020
• Subjects: Algorithms; Datasets; Image segmentation; Motor vehicles; Occlusion; Pixels; Semantic segmentation; Semantics; Traffic; Traffic accidents; China
• ISSN: 1742-6588; 1742-6596
• DOI: 10.1088/1742-6596/1550/3/032071

China's road traffic environment has the characteristics of mutual occlusion between targets, common use of motor vehicles and non-motor vehicles, etc., causing the problem of reduced accuracy of semantic segmentation. KITTI, Cityscapes and other datasets cannot meet the specific needs of semantic segmentation of traffic environment images in China, so Apolloscape dataset is used. The traditional method uses a single evaluation standard, and lacks the consistency check of the image semantic segmentation results. It ignores the relationship between pixels and pixels, which easily causes misidentification and leads to traffic accidents. On the basis of traditional cross-comparison evaluation indicators, this paper adds a composite evaluation of traffic environment semantic segmentation, emphasizing the role between pixels, making the results more consistent. Furthermore, it is through using the algorithm in perceiving the three typical traffic environment for verifying effectiveness. As a result, multi-task learning method is effectively applied in the three environments to achieve similar performance without increasing the computational overhead.